Discussion PaperDeutsche BundesbankNo 19/2018

International trade andretail market performance and structure:theory and empirical evidence

Philipp Meinen(Deutsche Bundesbank and Aarhus University)

Horst Raff(Kiel University, Kiel Centre for Globalization, and CESifo)

Discussion Papers represent the authors‘ personal opinions and do notnecessarily reflect the views of the Deutsche Bundesbank or the Eurosystem.

Editorial Board:

Deutsche Bundesbank, Wilhelm-Epstein-Straße 14, 60431 Frankfurt am Main,

Postfach 10 06 02, 60006 Frankfurt am Main

Tel +49 69 9566-0

Please address all orders in writing to: Deutsche Bundesbank,

Press and Public Relations Division, at the above address or via fax +49 69 9566-3077

Internet http://www.bundesbank.de

Reproduction permitted only if source is stated.

ISBN 978–3–95729–464–7 (Printversion)

ISBN 978–3–95729–465–4 (Internetversion)

Daniel Foos

Thomas Kick

Malte Knüppel

Jochen Mankart

Christoph Memmel

Panagiota Tzamourani

Non-technical summary

Research Question

In recent decades, advanced economies have experienced a sharp increase in consumer

goods imports, leading to an unprecedented degree of import penetration across many

consumer goods industries. In the current paper, we aim to determine whether this growth

in consumer goods imports has contributed to observed changes in the performance of

retailers and the structure of retail markets. The existence of such a link is suggested

not only by the fact that most of these imports pass through the retail sector, but also

that retailers themselves have played an active part in the rapid expansion of trade in

consumer goods.

Contribution

To guide the empirical analysis, we construct a model featuring heterogeneous retailers

that endogenously decide whether to import and whether to operate as chains. The main

economic mechanism we want to explore in our analysis builds on economies of scale in

importing, which imply that only big retailers and retail chains can afford direct imports

so that they benefit more from trade cost reductions than small retailers. We test the

model’s predictions using detailed Danish microdata for the period 1999-2008. In doing

so, we consider model-implied adjustments at the firm level and local retail market level.

Results

The empirical analysis shows that importing retailers are larger, more profitable, and have

a higher propensity to have multiple shops than domestically sourcing firms. While this

is partly due to self-selection, we also present evidence for improved performance caused

by firms’ importing activities. Moreover, we find that retail imports are associated with

a higher exit probability of small retailers and greater local retail market concentration.

Overall, we find support for the model’s predictions and argue that the observed adjust-

ments may imply additional gains from trade absent from models lacking a distribution

sector.

Nichttechnische Zusammenfassung

Fragestellung

In den vergangenen Jahrzehnten kam es in den fortgeschrittenen Volkswirtschaften zu

einem starken Anstieg von Konsumguterimporten. Das vorliegende Papier geht der Frage

nach, ob dieses Importwachstum zu den beobachteten Veranderungen in der Performance

von Einzelhandlern und der Struktur von Einzelhandelssektoren beigetragen hat. Ein sol-

cher Zusammenhang ist nicht nur naheliegend, weil die meisten dieser Einfuhren durch den

Einzelhandelssektor geleitet werden, sondern auch, weil Einzelhandler selbst eine aktive

Rolle bei der Ausweitung der Importe gespielt haben.

Beitrag

Die vorliegende Studie entwickelt ein Modell mit heterogenen Einzelhandlern, die daruber

entscheiden konnen, ob sie direkt importieren und ob sie als Handelskette agieren. Der

Hauptmechanismus, der in dem Modellrahmen untersucht wird, beruht auf Skalenertragen

im Zusammenhang mit der Importtatigkeit. Skalenertrage bedeuten, dass nur große Ein-

zelhandler und Handelsketten in der Lage sind zu importieren und diese daher starker

von einer Handelsliberalisierung profitieren als kleinere Einzelhandler. Das Modell impli-

ziert eine Reihe von Hypothesen, die anhand detaillierter danischer Mikrodaten fur den

Zeitraum 1999-2008 untersucht werden. Dabei werden die Folgen einer verstarkten Han-

delsintegration sowohl auf der Unternehmensebene als auch auf der (lokalen) Marktebene

betrachtet.

Ergebnisse

Die empirische Analyse zeigt, dass importierende Einzelhandler hohere Umsatze haben,

profitabler sind und mit hoherer Wahrscheinlichkeit mehrere Geschafte betreiben als Ein-

zelhandler, die ihre Waren ausschließlich vom heimischen Markt beziehungsweise uber

Zwischenhandler beziehen. Obwohl dies zum Teil auf Selbstselektionseffekte zuruckzufuhren

ist, legen die Ergebnisse der Studie auch nahe, dass Importtatigkeit die Performance von

Einzelhandlern verbessern kann. Daruber hinaus geben die Schatzungen Hinweise darauf,

dass sektorale Einzelhandelsimporte mit einer erhohten Wahrscheinlichkeit des Markt-

austritts kleiner Einzelhandler und einer großeren Marktkonzentration in Verbindung ste-

hen. Insgesamt weist die empirische Analyse auf die Plausibilitat des im theoretischen

Modell betrachteten Mechanismus hin. Weitergehende Uberlegungen zu den Wohlfahrts-

implikationen deuten auf zusatzliche Vorteile des internationalen Handels hin, die bisher

so nicht betrachtet worden sind.

Deutsche Bundesbank Discussion Paper No 19/2018

International Trade and Retail Market Performance

and Structure: Theory and Empirical Evidence∗

Philipp MeinenDeutsche Bundesbank, Aarhus University

Horst RaffKiel University, Kiel Centre for Globalization, and CESifo

Abstract

Based on a theoretical model featuring heterogeneous retailers that may sourceglobally and operate as chains, we derive a number of hypotheses that link tradeintegration to retail firm performance and to the structure of retail markets. Weempirically test these predictions using Danish microdata for the period 1999 to2008. We find that importing retailers are larger, more profitable, and have a higherpropensity to have multiple shops than domestically sourcing firms. While thisis partly due to self-selection, we also present evidence for improved performancecaused by firms’ importing activities. Moreover, we find that retail imports areassociated with a higher exit probability of small retailers and greater local retailmarket concentration. Overall, we obtain support for the model’s predictions andargue that the observed adjustments may imply additional gains from trade absentfrom models lacking a distribution sector.

Keywords: international trade, consumer goods, retailing, retail chains

JEL classification: F12, L11

∗Contact address: Philipp Meinen, Deutsche Bundesbank, Wilhelm-Epstein-Strasse 14, 60431 Frank-furt am Main; philipp.meinen@bundesbank.de. Horst Raff, University of Kiel, Department of Economics,Wilhelm-Seelig-Platz 1, 24118 Kiel; raff@econ-theory.uni-kiel.de. We are grateful to the Tuborg ResearchCentre for Globalisation and Firms (TGF) and its director, Philipp Schroder, for granting us access to thedata used in this paper. The data builds on anonymized micro data sets owned by Statistics Denmark.We would like to thank Emek Basker, Matilde Bombardini, Peter Egger, Carsten Eckel, Stefan Goldbach,Jan Marcus, Arne Nagengast and audiences at the CESifo Global Economics Conference, the EuropeanEconomic Association Conference, the ETSG Conference, the Aarhus-Kiel Workshop, the Gottingen In-ternational Economics Workshop and various seminars for very helpful comments and suggestions. Theviews expressed in this paper are those of the authors and do not necessarily coincide with the views ofthe Deutsche Bundesbank or the Eurosystem.

1 Introduction

In recent decades advanced economies have experienced a sharp increase in consumergoods imports, leading to an unprecedented degree of import penetration across manyconsumer goods industries.1 In the current paper, we aim to determine whether thisgrowth of consumer goods imports has contributed to observed changes in the performanceof retailers and the structure of retail markets. Specifically, we want to know how greaterimports – more precisely the fall in trade costs behind the import growth – affect the sales,markups, and profits of retailers, and what greater imports imply for the retail industry asa whole, including market concentration, firm exit rates, and the consolidation of retailersinto retail chains. The existence of such a link is suggested not only by the fact that mostof these imports pass through the retail sector, but also that retailers themselves haveplayed an active part in the rapid expansion of trade in consumer goods.2

Finding answers to our questions matters from both a positive and a normative per-spective. From a positive perspective, understanding what drives changes in retailing,including the growth of big retailers and retail chains and the exit of small retailers, mat-ters not only because retailing is a big sector, accounting for around 10% of employmentin many countries, but also because these changes have been a major source of retail pro-ductivity growth and, more importantly, aggregate productivity growth in industrializedcountries since the mid-1990s (see, for instance, Triplett and Bosworth, 2004). Our pa-per provides empirical evidence that the increase in consumer goods imports contributedsignificantly to the observed changes in retailing.

From a normative perspective, understanding what impact trade has on retailing mat-ters, because, with retail costs and markups often accounting for 30 to 50% of retail prices(Campa and Goldberg, 2006), any increase in retail productivity or reduction in markupspotentially has big welfare effects. What is more, these welfare gains would come ontop of the gains typically associated with international trade, such as gains from greaterproduct variety or production efficiency. We examine the source of these welfare gainsand explain why the trade-induced changes in retail performance and structure that wedetect in the data are consistent with an improvement in social welfare.

We base our analysis on Danish microdata for the period 1999 to 2008. These data arewell suited for two reasons. First, developments in consumer goods trade and retailingin Denmark mirror those in other industrialized countries; the Danish data may thusoffer a glimpse into economic mechanisms that may operate in a broader set of countries.According to the quantity index published by Statistics Denmark, imports of goods forhousehold consumption increased by 75% in real terms between 1999 and 2008, compared

1 For instance, import penetration in the apparel market is around 94% in the United States, 95% inGermany and the UK, 85% in France, 65% in Italy, and 55% in Spain. Import penetration in the USfootwear market is at 85% (Gereffi and Frederick, 2010). Average import penetration in textiles, clothingand footwear in the OECD stands at 59.4% (Nordas, 2008).

2 In the United States, retailers (including firms that engage in both retailing and wholesaling) repre-sent 14% of all US importing firms and account for 9% of the total value of imports (not just consumergoods) (Bernard et al., 2010a). Direct imports by retailers account for 31% of total US imports in textilesand clothing (HS 50-63), and for 34% of total US imports of footwear (HS 64-67) (Gereffi and Frederick,2010; Bernard et al., 2010b). In Canada the top 5% of importing retailers account for 76.3% of totalCanadian imports of clothing, shoes, jewellery, luggage and leather goods, and for 68.2% of all Canadianimports of electronics and appliances (Raff and Schmitt, 2016a).

1

to an increase in total real imports of 50%. In nominal terms, consumer goods importsrose by around 85%, with certain types of goods experiencing growth rates well above100% over this period.3 Import shares rose across many consumer goods industries, moststrongly in footwear (HS 64-67) from 65% in 1999 to 96% in 2008. Consumer goods thusspearheaded the overall increase in imports from 33% of GDP in 1999 to 51% in 2008.4

Retailers in Denmark, as elsewhere, have become major direct importers of consumergoods. In 2008, retailers in Denmark, more precisely firms that report retailing as theirmain activity, accounted for around 12% of all firms importing consumer goods and 14%of total consumer goods import values, with retailers’ shares of imports being much higherin some of the big consumer goods industries, such as furniture, toys and miscellaneousmanufactured articles (33%).

Second, the Danish data contain enough information at the firm level to allow us todeal with several conceptual and empirical challenges that are specific to the study ofretail markets. One such challenge consists of identifying the geographic scope of retailmarkets. Competition among retailers tends to be localized simply because consumerstypically do not travel long distances to go shopping (Maican and Orth, 2017). Thissuggests, among other things, that retail market concentration should be measured at thelocal level. In our empirical analysis we use information on municipalities to identify localretail markets, and we compute local market concentration based on the shops in eachlocal market. Another closely related problem stems from the fact that some retailers,typically the bigger ones, are organized as chains operating shops in several local retailmarkets.5 In the data, we are able to identify the shops belonging to each retail firm, andwe define retailers with more than one shop as retail chains. This allows us to study theconsolidation of retailers into chains and, more precisely, to determine whether greaterconsumer goods imports have contributed to this consolidation.

The main economic mechanism we want to explore in our analysis builds on economiesof scale in direct importing, which imply that only big retailers can afford direct imports.Smaller retailers, if they have access to imports at all, have to rely on more expensiveindirect imports via intermediaries.6 As a consequence, big retailers benefit more from

3 These goods include, for instance, imports of consumer goods related to animal or vegetable fats(HS15), electrical equipment (HS84-85), transport equipment (HS86-89), and furniture and toys (HS94-98). Consumer goods are defined based on the BEC classification, which can be merged to HS six-digitproduct codes. Specifically, BEC product codes 112 (food and beverages, primary, mainly for householdconsumption), 122 (food and beverages, processed, mainly for household consumption), 522 (transportequipment, non-industrial), 61 (consumer goods not elsewhere specified, durable), 62 (consumer goodsnot elsewhere specified, semi-durable), and 63 (consumer goods not elsewhere specified, non-durable) areconsidered to be consumer goods.

4 As a comparison, in euro area countries this ratio rose from 30% to 39% during that period.5 As already mentioned above, the consolidation of retailers into chains is an important phenomenon

that we observe in Denmark and elsewhere. In Denmark, the number of chains increased from around700 in 1999 to more than 1,200 in 2008. This is in line with observations in the United States, wherelarge retail chains (with at least 100 establishments) doubled their share of US retail sales from 18.6% in1967 to 36.9% in 1997 (Jarmin et al., 2009).

6 Significant economies of scale in direct importing activities of retailers have recently been docu-mented by Holmes and Singer (2017) for the United States. They show that big retailers like Walmart,Target and Costco, by nature of their large import volumes, face much lower “indivisibility” costs whenimporting goods via containers than small retailers that typically rely on intermediaries, including freightforwarders to deal with shipping companies, and on logistics firms to manage consolidation of shipmentsinto containers. These “indivisibility” costs consist of the cost of unused container space, the cost as-

2

a fall in trade costs than small retailers, which has consequences for the performanceof individual retailers and ultimately for retail market structure and productivity. Weexplore these consequences in a theoretical model of the retail industry, and then take thetheoretical hypotheses to the data. In our empirical analysis, we tend to find support forthe theoretical predictions. Firm-level regressions indicate that, after controlling for firmcharacteristics, sales, profits, and markups of directly importing retailers exceed those ofdomestically sourcing firms on average by 20%, 27%, and 3%, respectively. While theseresults could be solely due to larger and more profitable firms self-selecting into importing,further estimations based on propensity score matching techniques suggest that theseperformance differences are at least partly due to importing. Indeed, we estimate thatimport starters that were similar in observable characteristics to domestically sourcingfirms before beginning to import exhibit 8% greater sales, 6% greater profits, and 2%greater markups in the year of import initiation. The estimated effects turn out to bequite persistent. For instance, we estimate that, after three years, cumulative sales ofimport starters are on average over 30% higher than for comparable non-importers. Wealso find evidence that direct importing may increase the probability of an import starterbecoming a chain. This effect occurs two years after import initiation, with the probabilityrising by up to 16%.

Turning to the effects on retail market structure, we observe that an increase inindustry-level direct imports is associated with higher exit probabilities of small retailersand, through a decrease in the overall number of retailers, with greater market concen-tration at the local level. Including a proxy for indirect imports of consumer goods bywholesalers and other non-retail firms in our regressions provides additional evidence thatthese market structure effects are indeed driven by direct retail imports. Taken together,our results offer support for the direct importing mechanism outlined by our model, andindicate a non-negligible role of direct imports of consumer goods for the performanceand structure of the retail industry.

The current paper builds on the literature on the effects of trade on retail markets(recently surveyed by Raff and Schmitt, 2016a). Our theoretical model extends Raff andSchmitt (2012) by endogenizing the number of shops operated by a retailer and thusintroducing the decision of whether to become a chain. The shop margin arises from fixedcosts per shop as in the model of Basker and Van (2010a). In Basker and Van’s model,cheaper imports allow a single chain retailer to expand the number of shops, forcingsmaller, single-shop competitors to exit. Our main modelling contribution can indeedbe interpreted as putting their mechanism of adjustment along the shop margin into anindustry equilibrium model of retailing with heterogeneous firms and endogenous marketstructure, as is in Raff and Schmitt (2012).7

sociated with consolidation of different shipments into a single container, and the cost associated withdistorting the shipment size to fit in a standard container.

Economies of scale in direct importing also stem from fixed costs of importing, including the cost ofidentifying and dealing with suppliers, maintaining overseas buying offices, as well as large investmentsin logistics, inventory management and information technology. A survey of German, Swiss and Austrianretailers by Zentes et al. (2007) suggests that these costs are very high in practice and only borne by bigretailers, whereas small retailers import at most indirectly via wholesalers.

7 An alternative theoretical approach is suggested by Eckel (2009). In his paper, retail market con-centration is also driven by imports. But it does not come from an increase in the volume of consumergoods trade, but from a rise in the number of varieties available on the world market. As retailers expand

3

Our empirical analysis is most closely related to Basker and Van (2010b), who testwhether retail industries experiencing an increase in market concentration, measured bysales growth of the four largest firms, are more likely to sell goods exhibiting an increase inimports. Using US import data at the product level, they find that, indeed, product-levelimports grow especially fast when these products are disproportionately sold by indus-tries exhibiting high sales growth of the top four firms; this effect is particularly strongfor imports from China and other less-developed countries (LDCs). In a counterfactualexercise, they attribute half of the growth of imports from China and other LDCs toincreased retail industry concentration.

Like Basker and Van, we find a significant correlation between imports and retailmarket concentration. By our estimates, a 1% increase in industry-level direct importsraises the Herfindahl index of local retail market concentration, ceteris paribus, by upto 0.042%. This is not a small effect, considering that average direct imports in theestimation sample more than doubled during the sample period and that the Herfindahlindex is at a relatively high level initially.8

Since we have access to both firm-level and shop-level data, we can examine the differ-ential performance of direct importers and non-importers with respect to sales, markups,profits, the probability to becoming a chain, and the exit probability, and thus go deeperthan Basker and Van into trying to uncover possible economic mechanisms driving thecorrelation between imports and retail market concentration. For instance, we can showthat the sales growth of large, importing firms, which is taken as the exogenous variableby Basker and Van, can be partly attributed to import activity.

By showing how trade can help explain observed changes in retail market performanceand structure, our paper complements studies that have focused on technology adoptionas a driver of retail market changes.9 Our paper can further be seen as complementingthe study by Holmes and Singer (2017), which examines how retailers respond to greaterdirect imports of consumer goods by altering the geographic structure of their importdistribution.10

Finally, given the focus of international trade research on the manufacturing sectorand the dearth of studies on the service sector in general and on retailing in particular,we want to point out how our paper compares to the more familiar research on the effectsof trade exposure on manufacturing. An obvious difference regarding the effect on firm

their assortment to include more imported varieties, their fixed costs rise and they have to have greatersales, markups and operating profits to cover these fixed costs.

8 We should point out that a potentially attractive feature of Basker and Van’s data is that theycapture both direct and indirect imports. However, as already mentioned, we can also create a proxy forindirect imports and find that it generally does not have a statistically significant effect on retail marketconcentration.

9 See, for instance, Holmes (2001), Basker et al. (2012), and Lagakos (2016). Foster et al. (2016)provide a recent survey of the literature, looking mostly at technology-driven changes in retailing. Anotherpotential driver of these changes is entry by multinational retailers (see Atkin et al., 2018).

10 Our paper is also indirectly related to the empirical literature on exchange rate pass-through intoretail prices (see, for instance, Antoniades and Zaniboni, 2016). This literature has recognized that theretail margin (i.e. costs of retailing and retailer markups) is a major factor in explaining why changesin import prices are only incompletely passed through to retail prices. Antoniades and Zaniboni, inparticular, find that pass-through varies systematically by retailer size. We differ from this (short-run-oriented) literature in that we study the effects of imports in the long term, where markups and retailmarket structure are endogenously determined.

4

performance is that retail services are traditionally non-tradable, so that retailers typicallydo not face import competition and do not export. In fact, we argue that retailers typicallyeven face very little interregional competition. At least in our sample period, retail salesnot in stores (i.e. over the internet, by mail order, etc.), where import competition andexports are most likely to arise, are very low, and we exclude them from the analysis.The trade exposure of retailers in our study therefore mostly comes from the goods theyimport. The performance effects we measure for retailers are therefore best compared tothe effects on manufacturers stemming from the import of intermediate goods (see, forinstance, the recent studies of De Loecker et al. (2016) and Brandt et al. (2017) on themarkup and productivity effects from reductions in input tariffs).

Regarding market structure, the non-tradability of retail services has the distinctadvantage that we can, at least with a greater degree of confidence than in the caseof manufacturing, measure market concentration, simply because market demarcation iseasier. Retailing may therefore even offer a better opportunity than manufacturing forstudying the market structure effects of importing.

The rest of the paper is organized as follows. In the next section, we present ourtheoretical framework and derive testable hypotheses about the effect of trade on retailmarkets. In section 3, we describe the data, discuss sample selection, and provide descrip-tive evidence. In section 4, we outline the empirical approaches for testing each hypothesisand present the corresponding results. Section 5 concludes by discussing what our resultsimply for social welfare and, more specifically, for the gains from trade. Proofs, summarystatistics, and additional estimation results are presented in the Appendix.

2 Theoretical framework

In this section, we present a simple partial equilibrium model of a retail industry, inwhich retail firms differ in terms of their productivity and in which the number of retailfirms and the number of shops each firm operates across different local retail markets areendogenously determined. We use this model to formulate testable hypotheses about howa reduction in trade costs affects retailer performance and the structure of retail markets.

2.1 The model

Consider a country divided into R local retail markets. Each active retail firm has atleast one and at most R shops, i.e. no more than one per local retail market. A firm withmore than one shop, and hence operating in more than one local retail market, is calleda retail chain.

Consumer preferences in each local market follow the ‘random preference Hotelling’framework of Innes (2006).11 We therefore assume that, in local market r = 1, ..., R, thereis a measure Lr of consumers uniformly distributed around a circle of unit circumference.Each consumer visits a shop to purchase one unit of an aggregate consumption good

11 An alternative specification, more familiar to trade economists, would be to assume linear quadraticpreferences as in Melitz and Ottaviano (2008). See also Raff and Schmitt (2012) for the use of thesepreferences. Which preference specification is used makes no difference to our hypotheses. We use thecurrent specification to obtain a better microfoundation for consumer demand so that it becomes easierto discuss the possible welfare consequences of our results.

5

and faces a linear transport cost τ per unit of distance to visit the shop. There are nrshops located symmetrically on the circle. Preferences are random in the sense that eachordering of the nr shops among the nr locations is possible and occurs with the samerelative frequency. A simple interpretation of these preferences is that, when choosingwhere to shop, a consumer chooses only between the two nearest shops, as in a standardaddress model, but the identity of the two shops depends on an unobserved, randomevent.12

We assume, for simplicity, that local markets are symmetric, and therefore drop thesubscript identifying the local market. The aggregate demand qi faced by shop i can thenbe shown to be linear in prices, where qi(pi) = L

n− L

τpi + 1

n−1Lτ

∑nh=1,h6=i ph. We make the

additional assumption that n is large enough so that demand can be approximated by:

qi(pi) =L

n− L

τpi +

L

τp, (1)

where p is the average local retail price.Now consider the level of the firm. To enter the market, a firm incurs a sunk cost

FE, which includes the cost of setting up one shop. After entering, each firm learns itsmarginal retail cost c (or productivity 1/c), which applies to all the shops it operates.The distribution of c is denoted by G(c) with support on [0, cM ]. We let productivityfollow a Pareto distribution, so that the cumulative distribution function for c is:

G(c) =

(c

cM

)k, (2)

with k ≥ 1.Imports are decided at the firm level. Economies of scale in importing arise because

direct importing involves a fixed cost FI , which includes the cost of maintaining over-seas buying offices and cooperating with foreign partners to source goods, the cost ofinformation technology needed to manage complex international sourcing operations, etc.Upon entry, each firm has to decide whether to source the aggregate consumption gooddomestically (which may include imports sourced indirectly through wholesalers) or torely on direct imports.13 Purchasing the good domestically is associated with a cost of wper unit. If the firm relies on direct imports, the unit cost of the good is t < w, where talso includes transport and other costs associated with international trade; for simplicity,we refer to t below as the trade cost.

Another decision taken upon entry at the firm level concerns whether to become achain and, more precisely, how many shops to operate. We assume that each additionalshop beyond the one set up when the firm enters the market involves a fixed cost FS.

The retail industry is monopolistically competitive. At the local market level, thisimplies that each shop, respectively firm, takes the number of active shops in the market,n, and the average local retail price, p, as given when setting its price. A shop i belonging

12 Each shop therefore competes not just with its neighbors on the circle, but with all other local shops.This assumption is important because it allows us to derive a free-entry equilibrium in which shops havedifferent costs, which would not be feasible in a standard address model.

13 This assumption is not restrictive, since the aggregate consumption good can be interpreted as acomposite good consisting of purely domestic goods and goods that may be either imported directly orsourced domestically. See Raff and Schmitt (2012) for details.

6

to a firm with marginal cost c obtains an operating profit of:

(pi − c− x)qi(pi), (3)

where x = w if the firm relies on domestic sourcing, and x = t if the firm is a directimporter. Using superscript D to indicate domestic sourcing, I to indicate importing, anddefining cD ≡ τ/n+ p−w, a shop with marginal cost c has the following profit-maximizingprices and sales when goods are sourced domestically or imported, respectively:

pD(c) = c+ w +1

2(cD − c) ; (4)

pI(c) = c+ t+1

2(cD − c+ w − t) ; (5)

qD(c) =L

2τ(cD − c) ; (6)

qI(c) =L

2τ(cD − c+ w − t) . (7)

Hence cD represents the marginal cost at which a shop belonging to a firm sourcingdomestically optimally chooses zero sales and thus to be inactive, i.e. qD(cD) = 0.

A shop with marginal cost c earns an operating profit equal to:

πD(c) =L

4τ(cD − c)2 , or (8)

πI(c) =L

4τ(cD − c+ w − t)2 , (9)

depending on whether the firm relies on domestic sourcing or importing. Only shopsbelonging to firms with marginal costs less than or equal to cD will remain active becauseonly they will be able to cover their marginal cost.

The decisions of whether to import and how many shops to operate are obviouslyinterdependent: a chain, for instance, is able to spread the fixed cost of importing acrossits shops and thus more likely to import than a single-shop firm. To avoid confrontingthe reader with a plethora of cases involving different shop-importing combinations, wemake several assumptions, described below, to ensure that only firms that are productiveenough to import will want to operate more than one shop. As we will show below, thiscaptures the empirically most relevant case.

The firm that is just indifferent between domestic sourcing and direct importing isthen a single-shop firm for which πD(c) = πI(c) − FI . This condition defines a criticalvalue of the marginal cost cI :

cI = cD +(w − t)

2− 2τFIL(w − t)

, (10)

such that firms with c ≤ cI prefer importing and firms with c > cI prefer domesticsourcing. We assume that cI < cD so that the least efficient active firms engage indomestic sourcing; sufficient conditions are given in the Appendix.

How many shops will an importing firm operate? Since, by assumption, local mar-kets are symmetric and the fixed cost of each additional shop is constant, the answer is

7

straightforward, namely either a single shop or R shops – one in each local market. Inparticular, an importing firm is indifferent as to whether to open an additional shop ifπI(c) = FS. Denote the marginal cost at which this condition is satisfied by cS. Sincewe require that cS < cI so that, as already discussed above, only importing firms becomechains, the only admissible solution is:

cS = cD + w − t− 2

√τFSL

. (11)

Sufficient conditions to ensure 0 < cS < cI are given in the Appendix.The three cut-off values of the marginal cost, cD, cI and cS, define four categories of

firms. The most productive firms, i.e. those with marginal cost c ≤ cS, import directlyand operate as chains; firms whose marginal cost is in the interval cS < c ≤ cI importdirectly but operate only a single shop; firms in the interval cI < c ≤ cD are also single-shop firms but do not import directly; and firms with high marginal costs (c > cD) areinactive.

To close the model, consider the entry decision firms face before observing theirmarginal costs. Firms enter the retail industry until their expected profits are zero:

∫ cS

0

[RπI(c)− (R− 1)FS − FI

]dG(c)+

∫ cI

cS

[πI(c)− FI

]dG(c)+

∫ cD

cI

πD(c)dG(c)−FE = 0.

(12)

2.2 Testable hypotheses

In this subsection, we examine the comparative statics of the model with regard to changesin the trade cost t and formulate corresponding hypotheses; all proofs are in the Appendix.We start by checking how the cut-off value cD changes with t. Applying the implicitfunction theorem to the zero-profit condition (12) allows us to show that dcD/dt > 0.This implies the following hypothesis:

Hypothesis 1 A reduction in the trade cost forces the least efficient firms to becomeinactive.

The intuition for this effect is straightforward. A reduction in the trade cost, ceterisparibus, raises the profits of direct importers whether they are single-shop firms or chains.To hold expected profits at zero, cD has to decrease so as to lower the probability of beingan active firm.

The sign of dcI/dt can now be obtained from (10). We find that dcI/dt < 0, whichimplies that a trade cost reduction induces some firms that were previously not productiveenough to afford the associated fixed cost to switch to direct importing. In particular, itis the most productive non-importers that self-select into importing directly. Hence, wemay state:

Hypothesis 2 A reduction in the trade cost induces the most productive non-importersto switch to importing directly.

8

What does a fall in the trade cost imply for firm performance? Holding fixed thenumber of shops a firm operates, we can show that the sales of direct importers rise,and so do markups because, as can be seen in (5), importers pass only half of what theysave on trade costs on to consumers. In fact, this direct effect on markups dominates theindirect effect stemming from an increase in the price elasticity of demand induced by alower cD. Hence the profits per shop of importers rise. Firms that do not import directlyexperience increased pressure on their markups from the rise in the price elasticity ofdemand, but obviously do not enjoy any offsetting cost savings. This forces them to cuttheir markups and sales, which leads to lower profits.14 These effects can be summarizedas follows:

Hypothesis 3 A reduction in the trade cost (i) raises the sales, markups and profits pershop of firms that engage in direct imports; and (ii) lowers the sales, markups andprofits of firms that do not import directly.

Hypothesis 3 thus points to starkly different performance of importing and non-importing firms following a trade cost reduction. A potential strategy for testing thishypothesis is to focus on firms that switch to direct importing. While Hypothesis 2 sug-gests that part of the performance difference between import starters and non-importersis due to self-selection of larger, higher markup, and therefore more profitable firms, intoimporting, Hypothesis 3 indicates that import starters should experience an additionalboost to their performance caused by their access to cheaper direct imports.

The next hypothesis is associated with the sign of dcS/dt and thus with the effect of atrade cost reduction on the consolidation of retailers into chains. As can be seen from (11),a marginal change in t has two effects on cS. The direct effect is negative: a reduction inthe trade cost raises the profit that a direct importer can earn in each local market andthereby allows some firms to cover the fixed costs associated with operating additionalshops and thus becoming a chain that were previously unable to do so; therefore cS rises.The indirect effect comes from the zero-profit condition and is positive: a reduction in treduces cD and thus lowers the probability of being active, which tends to lower cS. Wecan show that the direct effect dominates so that:

Hypothesis 4 A reduction in the trade cost induces the most productive single-shop firmsto add shops and become chains.

Next, we examine how a reduction in the trade cost, by changing the marginal-costcut-offs and the performance of individual firms, affects the aggregate performance ofthe retail industry and the structure of local retail markets. A simple inverse measureof average retail productivity can be computed as the mean of marginal costs of activefirms:

c =1

G(cD)

∫ cD

0

cdG(c) =k

k + 1cD. (13)

Clearly, trade increases this measure of retail productivity by forcing the least efficientretailers to become inactive.

14 These firms are single-shop firms by assumption. But obviously these results would hold at the shoplevel, even if the firms were chains.

9

The average markup of active firms, µ, can be computed as:

µ =1

G(cD)

(∫ cI

0

[pI(c)− c− t

]dG(c) +

∫ cD

cI

[pD(c)− c− w

]dG(c)

)(14)

=(w − t)

2

ckIckD

+cD

2 (k + 1). (15)

The effect of a marginal decrease in the trade cost on the average markup can be decom-posed as follows:

dµ

dt= −1

2

ckIckD− k(w − t)

2

ckIckD

[1

cD

dcDdt− 1

cI

dcIdt

]+

1

2(k + 1)

dcDdt

. (16)

The first term in (16) gives the direct effect: importing firms pass only half of the tradecost reduction on to consumers, the other half goes to raising their markup. This hasto be weighted by the probability that the firm is an importer given that it is active,which is equal to ckI/c

kD. The second term reflects the fact that a trade cost reduction

tends to raise the average markup by changing the distribution of active firms in favorof higher-markup, importing firms. Specifically, a decrease in the trade cost increasesthe probability that the firm is an importer given that it is active (by reducing cD andincreasing cI). The third term represents an effect that goes in the opposite direction:a decrease in t, by reducing cD, raises the price elasticity of demand for all firms, whichtends to lower the average markup.

Whether a decrease in the trade cost increases or decreases the average markup is thusambiguous, and depends not least on the (endogenous) share of importers in the industry,here captured by ckI/c

kD. We may therefore state that:

Hypothesis 5 A reduction in the trade cost raises average retail productivity, but has anambiguous effect on the average markup.

From the cut-off cD we can compute the number of active shops in a local market as:

n =τ

(cD + w − p). (17)

This number reacts to a reduction of t in two ways. First, a lower t reduces the averageretail price p. This price effect tends to reduce the number of active shops. Second, adecrease in t reduces cD. This selection effect means that firms and the shops they operatebecome more efficient on average. This tends to increase the number of active shops. Thesign of dn

dtis therefore generally ambiguous. The price effect dominates if the fixed cost of

importing is not too great. Hence we can formulate that:

Hypothesis 6 A reduction in the trade cost lowers the number of shops in a local marketif the fixed cost of importing is sufficiently small.

Notice, however, that even if a fall in the trade cost decreases the number of shops,this does not necessarily imply an increase in market concentration at the local level. This

10

can best be seen when writing the Herfindahl index (H) of local market concentration asfollows:

H =1 + σ2

q/q2

n, (18)

where q denotes average sales per shop and σ2q is the variance of sales (see Waterson, 1984).

Thus, market concentration in a setting in which firms differ in their marginal costs, andthe shops they operate therefore differ in sales, is positively related to the coefficient ofvariation of retail sales, σq/q. That is, we have to check how lower trade costs affect thedistribution of sales across shops. For the average sales per shop and the variance of saleswe obtain:

Hypothesis 7 A reduction in the trade cost raises the mean and reduces the variance ofshop-level sales if the fixed cost of importing is sufficiently small.

The effect of a lower trade cost on the Herfindahl index is therefore generally ambigu-ous, simply because a fall in the number of shops may be accompanied by a fall in thecoefficient of variation of local retail sales. The question of how a lower trade cost affectslocal market concentration in retailing can therefore only be answered empirically.

3 Data

3.1 Data sources

For our empirical analysis, we use data on retail firms present in Denmark between 1999and 2008.15 Specifically, we make use of three data sets available from Statistics Denmark.The first data set, FIRM (“Generel Firmastatistik”), covers the population of firms activein Denmark and contains information on industry affiliation, number of employees, andother firm characteristics, such as turnover, value added, profits, and fixed and totalassets.16

The second data set, called UHDI (“Udenrigshandel diskretioneret”), provides infor-mation on individual firms’ export and import activities at a detailed product level and bypartner country. The data fall into two categories: Intrastat (for trade among EU mem-ber states) and Extrastat (for trade with countries outside the EU). Extrastat data comefrom custom forms and tax authorities and cover nearly all trade, while Intrastat data areself-reported figures by Danish firms that exceed certain export and import thresholds setby the EU.17

The third data set, IDAS (“IDA arbejdsstede”), contains information at the levelof the branch which, in the case of retail firms, is usually a shop. The data set includesinformation about the location of a branch, its industry classification as well as data aboutthe individuals working there. We use this information to identify local retail markets and

15 In order to accommodate a lag structure further described below, we also make use of informationfor the years 1997 and 1998.

16 We deflate firm-level domestic sales, value added and wages using the consumer price index with2000 as the base year.

17 For instance, in 2002, the thresholds were DKK 2.5 million for exports and DKK 1.5 million forimports. The thresholds are set each year for imports and exports separately in order to ensure coverageof 95% and 97% for imports and exports, respectively.

11

to compute Herfindahl indices for these markets. We also use it to compute the numberof shops by retail firm so that we can investigate potential adjustments of retail firms totrade along the shop margin.

Thanks to a unique firm identifier, we can merge the information present in each of thedata sets mentioned. Whereas FIRM includes the population of active firms and UHDIcontains information on international trading activities of firms in Denmark subject torather small reporting thresholds, the data set IDAS does not include all firms present inFIRM. In the next subsection, we describe in more detail how we constructed the datasets used for the empirical analysis.

3.2 Sample selection

Based on the industry affiliation documented in FIRM, we identify all firms that reportretailing as their main activity.18 Statistics Denmark reports the industry affiliation at thesix-digit level, where the first four digits correspond to the NACE four-digit classification.Generally, the NACE four-digit level appears most appropriate to define a retail industry.19

However, in case of the sector “other retail sale in specialized stores” (5248), the four-digitclassification is too broad, as it masks the large heterogeneity of retailer firms belongingto this industry (e.g. jewelry, sports equipment, toys, bicycles, electronics). Hence, forthis industry we use a more detailed industry definition.

We drop a few retail industries from our sample where the economic mechanismswe want to examine are likely to be absent, e.g. because direct importing is ratherinfrequent, which is the case for NACE rev. 1 three-digit industry 522 (retail sale of food,beverages and tobacco in specialized stores), 525 (retail sale of second-hand goods instores), 526 (retail sale not in stores), and 527 (repair of personal and household goods),or because the industry is heavily regulated, which is the case for industry 523 (retail saleof pharmaceutical and medical goods, cosmetic and toilet articles).20 In total, we thusretain 24 retail industries in our analysis, which we list in Table A.3 in the Appendix.

The empirical analysis is conducted at two different levels of observation. At the firstlevel, the firm is the unit of analysis, and we investigate firms’ adjustments related todirect importing of consumer goods.21 At the second level, we examine how aggregateindices of retail market structure derived from the theoretical model respond to consumergoods imports. For this purpose, we require a retail market demarcation that is not only

18 We therefore do not consider firms with main activities in other industries but which potentiallyhave branches engaging in retailing.

19 For instance, the industry retail sale of clothing (four-digit sector 5242) is further broken down intothe six-digit sub-industries retail sale of ladies’ clothing, retail sale of men’s clothing, retail sale of men’sand ladies’ clothing, and retail sale of baby articles and children’s clothing. Hence, at least the formerthree sub-industries overlap, which makes it difficult, for example, to compute the Herfindahl index.

20 The number of excluded firms decreases over the sample period from more that 7,000 in 1999 to lessthan 6,000 in 2008; their share in total retail firms’ imports also decreases during that time from 9% to7%, while their share in total retail firms’ sales remains relatively stable at around 13%.

21 Note that we clean the data by dropping observations with implausible values for key variables(e.g. negative sales) and by removing outliers in terms of the labor productivity distribution, defined asobservations that deviate from median by more than five times the standard deviation. Moreover, weclean dependent variables such as sales, profits or markups by dropping observations if the year-on-yeargrowth rate of the respective variable deviates by more than five times the standard deviation from themedian.

12

based on industry affiliation, but also on the geographic scope of the market.We define the geographic scope of a retail market to be a municipality. Foged and Peri

(2015, p.12) find that, in Denmark, “municipalities are, even in the long run, rather self-contained labor markets.” If workers on average seek employment only within their ownmunicipality, it seems plausible to assume that this is also where they go shopping. Butthis raises three data-related issues. First, in 2007, a new demarcation of municipalitieswas introduced in Denmark that reduced the number of municipalities from more than 270to 98. Most of the old municipalities are linked to only one new municipality. However,for a few old municipality codes, there is no one-to-one correspondence to new codes.We deal with this issue by forming some larger regional groups in order to concord themunicipality codes over time. We end up with 85 regions that we use to define local retailmarkets.

Second, we have to deal with multi-shop firms. While we observe the location, thenumber of employees and the wage bill at the shop level, information on sales, for example,is only available at the firm level. In the case of multi-shop firms, we therefore need todistribute firm-level sales across shops based on a weighting scheme. Given the informationavailable in IDAS, we use the share of a shop’s employment in a firm’s total employmentas weight.

Third, we face the problem that some active firms present in FIRM cannot be mergedto IDAS, implying that for these firms we do not observe information at the level of theshop. Hence, these firms are excluded from the analysis whenever shop-level informationis required, as is the case in the analysis of local retail market structure. This issue con-cerns in particular the two NACE four-digit industries 5211 (retail sale in non-specializedstores with food, beverages or tobacco predominating) and 5212 (other retail sale in non-specialized stores) for the years 1999 to 2001. We thus exclude these industries during thisperiod from the analysis of local retail market structure. For the remaining industries,we can merge firms that account for around 95% of retail sales in every year.

Tables A.1 and A.2 in the Appendix present summary statistics for variables used inour analysis.

3.3 Descriptive evidence

Table 1 presents a number of aggregate indicators by year for the retailers in our sample.Column (i) indicates that retailers accounted for an important share of firms importingconsumer goods and that this share remained relatively constant during the sample periodat around 12%.22 Retailers also mattered with respect to the intensive margin of consumergoods imports and this involvement increased over time. As indicated by column (ii), theshare of imports by retailers in total consumer goods imports increased from 9.2% in 1999to 13.8% in 2008. Indeed, for certain product categories, retailers accounted for a muchlarger share of consumer goods imports.23

22 Over the sample period, consumer goods on average account for almost 80% of total imports of ourretail firms.

23 For instance, in 2006, retailers accounted for 23% of imports of wood and articles of wood (HS44-46), 20% of imports of raw hides and skins, leather and articles thereof (HS 41-43), 33% of importsof furniture, toys, and miscellaneous manufacturing articles (HS 94-96), and 39% of imports of art andantiques (HS 97-99).

13

Between 1999 and 2008, we observe an increase in average retail sales per firm fromDKK 10 million to DKK 11.5 million (column iii). This increase is entirely due to di-rect importers whose average sales surged by almost 30%, while those of other retailersdecreased by more than 15% during that period. At the same time, the share of directimports of consumer goods in total retail sales more than doubled from 3.5% in 1999 to7.7% in 2008 (column iv). Moreover, average imports by directly importing retailers roseby more than 60% (column v). This latter development occurred despite a significant in-crease in the share of directly importing retailers which rose from 7.1% to 10.8% (columnvi).24

Finally, in order to get an idea about the degree of retail market concentration andhow it has changed over time, the last column of Table 1 presents the development of theaggregate Herfindahl index computed as the sales-weighted average of the indexes in localretail markets.25 The aggregate index remains relatively constant over time at around0.24.

Table 2 presents additional information about the role of retail chains during thesample period, where a chain is defined as a firm that has more than one shop.26 Firstof all, we see that the number of chains increased considerably during the sample period(by more than 70%). Moreover, we can observe that these firms are quite distinct since,despite constituting a relatively small group, they account for 64% of total retail turnoverin 2008 (up from 55% in 1999). These firms also play a crucial role when it comes toimporting; one in four chains imports directly, and, in total, chains account for around90 % of total direct imports by retailers. Furthermore, it is worth noting that there issubstantial heterogeneity across chains. While the median chain only has two shops, therealso is a small number of chains with well above 100 shops. Indeed, when only focusingon chains with at least ten shops, then these chains alone account for 50% of total retailturnover in 2008.27

Table 3 underlines these points by distinguishing retail firms according to their importand chain status. This results in a sorting in terms of average size which is quite plausiblealso with respect to our theoretical model. In particular, we find that importing chainretailers are by far the largest firms in terms of average sales and number of employees,followed by non-importing chains and importing single-shop firms. Moreover, we observethat average imports by chain retailers that import directly exceed those of single-shopimporters by a factor in excess of 30. Similarly, according to these numbers, the uncon-ditional probability of a chain store importing amounts to more than 25%, while thatof single-shop firms lies below 10%. Overall, these results thus suggest a relationshipbetween firm performance and import activity, which we will further analyze below.

24 Note that the total number of retail firms remained rather constant over time.25 The Herfindahl index is computed according to equation (18) for local retail markets. Note that

retail sectors 5211 and 5212 are excluded from these averages for the reason described in section 3.2.26 The information in this table is based on firms that are present both in FIRM and IDAS.27 Note that a firm with ten shops corresponds to the last decile of the shop distribution across all

chains.

14

4 Empirical analysis

The hypotheses derived from the theoretical model imply different performance of import-ing and non-importing firms in response to lower trade cost that translates into structuralchange in the industry and, in particular, in local retail markets. In this section, we em-pirically test the hypotheses. We first consider the predictions related to adjustments atthe firm level, and then focus on those related to structural changes at the industry andlocal market levels. For each hypothesis, we first lay out the empirical approach for takingit to the data, and then present the results. Further note that we use the approach by DeLoecker and Warzynski (2012) in order to obtain measures of markups that vary acrossfirms and time. We refer the reader to the Appendix for details about the estimationapproach and its empirical implementation.

Notice that, since we cannot measure trade cost directly, we test the hypotheses indi-rectly by assuming that the observed increase in import activity during the sample period,documented in Table 1, is at least in part driven by a trade cost reduction. Throughoutthe empirical analysis, we focus our attention on imports from countries outside the EU15for two main reasons. First, a significant trade cost reduction during the time period un-der investigation likely occurred through enhanced trade integration with respect to thesemarkets; we therefore use “trade cost reduction” and “trade integration” interchange-ably.28 Second, by excluding EU15 economies we largely avoid data issues related toreporting thresholds (see section 3.1) and potential biases that could arise from intra-firmtrade of multinational retailers headquartered in Sweden or Germany, for example.

4.1 Import activity and firm performance (H2 and H3)

According to Hypothesis 2, a decrease in the trade cost induces larger, higher markupand therefore more profitable retailers to self-select into importing. Moreover, Hypoth-esis 3 suggests that increased trade integration enhances these performance differences.Specifically, sales, profits, and markups of importing firms should increase, while those ofnon-importing firms should decrease.

Empirical approach (H2 and H3)

The first step of our empirical approach to test these predictions is to estimate importpremia regressions while controlling for other firm-level characteristics. Specifically, weestimate the following model:

yit = β0 + β1DIMPit + β2xi,t−1 + αst + αi + εit, (19)

where DIMPit is a dummy variable taking a value of one, if firm i imports consumergoods from at least one non-EU15 market in year t and it is zero for firms not importingfrom these markets;29 yit measures the logarithm of the outcome of interest, namely sales,

28 Examples of increased trade integration include the EU enlargement towards the East in 2004 andChina’s accession to the WTO in 2001.

29 Hence, the comparison group comprises firms that source domestically only and those that alsosource from EU15 markets.

15

profits, or markups, while xi,t−1 controls for firm size (number of employees);30 αst aresector-year dummies, implying that we are comparing importing and non-importing firmswithin sector-year pairs. In some specifications, we also include firm fixed effects, αi, inthe regressions so that the coefficient of interest, β1, is identified by firms experiencing achange in their import status.

Remember that both Hypothesis 2 and Hypothesis 3 suggest that importing firmsperform better than non-importers, implying a positive sign for β1. However, the expla-nations for a positive coefficient differ. According to Hypothesis 2, β1 should be positivebecause of self-selection, whereas Hypothesis 3 indicates that performance differences mayalso be caused by firms’ import activities due to access to cheaper products. While re-gressions based on equation (19) therefore provide first insights into differences betweenimporting and non-importing firms, these regressions provide no indication of causality.

We try to address this issue by taking our cue from studies investigating the rela-tionship between exporting/importing and the performance of manufacturing firms.31 Tothis end, we focus on retail firms that start importing. Specifically, we define a dummyvariable, STARTit, that equals one for firms that import in period t, but did not do soin periods t − 1 and t − 2, while it takes on zero for firms that have not imported dur-ing all three years. We then investigate the self-selection hypothesis by testing whetherimport starters perform better than non-importers before the import event occurs. Moreprecisely, we run the following regression:

yi,t−1 = β0 + β1STARTit + β2xi,t−1 + αst + εit. (20)

The differences compared to equation (19) are that we regress a lagged performancemeasure (yi,t−1) on a dummy variable indicating import starters (STARTit). Hence, β1

now informs us whether import starters already differ from non-importing firms one periodbefore the import event occurs, as suggested by Hypothesis 2.

Investigating Hypothesis 3, which suggests a causal relationship between importingand retail firm performance, is somewhat more involved. We try to identify this causaleffect by again focusing on import starters and then applying propensity score matching(PSM) in order to create a control group of firms that do not import but exhibit astatistically similar propensity to start doing so for each import starter.32 The variableStartit thus acts as a treatment dummy and we are interested in the difference betweeny1i and y0

i , where yi denotes again the performance of firm i, and the superscripts 1 and0 indicate the firm’s treatment status. More formally, we wish to compute the average

30 Note that controlling for the log of the number of employees implies that we restrict the analysis tofirms with at least one employee.

31 See, for instance, De Loecker (2007) or the survey by Wagner (2007). More recently, Smeets andWarzynksi (2013) present an investigation relying on similar methods that links both exporting andimporting to the performance of manufacturing firms.

32 When separately examining the performance of direct importers and non-importers, we would haveto use instruments for firm-level or industry-level imports in order to establish causal effects. Indeed,we experimented with some potential candidate variables (e.g. a country’s world export supply as inHummels et al., 2014), but the instruments generally turned out to be rather weak. Part of the problemappears to be specific to an analysis of retail firms because these firms significantly increased the number ofimported varieties during the period under investigation, while the instruments proposed in the literatureusually work for the intensive rather than the extensive margins of trade.

16

treatment effect on the treated defined as:

E[(y1 − y0)|Start = 1] = E[(y1)|Start = 1]− E[(y0)|Start = 1]. (21)

The fundamental evaluation problem inherent in equation (21) stems from the fact thatwe do not observe the counterfactual outcome E[(y0)|Start = 1], i.e. the expected perfor-mance of an import starter had it not switched to importing. A strategy for addressingthis problem involves the assumption that, conditional on a set of observable firm-levelcharacteristics that are unaffected by importing, potential outcomes are independent oftreatment. The latter is usually referred to as the conditional independence assumption,which implies that selection into treatment is driven by observable covariates. Intuitively,we try to construct a group of comparison firms that are as close as possible to the treatedfirms in terms of their propensity to start importing. To this end, we follow the insightspresented by Rosenbaum and Rubin (1983, 1985) and apply PSM.33 The first stage of thematching approach involves estimating the probability of starting to import:

P (Startit = 1) = Φ{zi,t−1}, (22)

where Φ(·) is the normal cumulative distribution function, indicating that we employa probit model; zi,t−1 contains firm-level control variables lagged by one year, i.e. oneperiod before import initiation. In particular, zi,t−1 includes measures of productivity(value added per employee) and size (number of employees and total assets), since bothfactors are commonly associated with firms engaged in international trade.34 Moreover,we condition on lagged sales growth and a firm’s wage share in total sales. The formervariable is meant to capture firms’ cyclical positions, while the latter variable is a roughmeasure of profitability. Note that we also include a quadratic productivity term to allowfor potential non-linearities. Finally, we add four-digit NACE industry and year dummiesto the probit regressions.

The second step of the matching approach involves the search for a control group thatis similar to the treated firms according to the propensity score estimated by the probitmodel. In particular, we apply radius matching with a tight caliper and impose commonsupport to ensure that the balancing property holds.35 In other words, for each treatedfirm, we search for a control group that consists of non-importing firms that differ in termsof the propensity score by no more than a pre-specified maximum distance (i.e. a caliperof 0.001).36 We can then compute the average treatment effect on the treated (ATT) as:

ATT =1

N

∑i

(y1i −

∑j∈Ci

wijy0j ), (23)

33 Caliendo and Kopeinig (2008) present an overview of propensity score matching techniques.34 Note that the productivity measure is prone to the caveat that, especially in smaller retail firms, the

owner (and potentially family members) may be an important part of a firm’s workforce without beingcounted as an employee. This variable therefore tends to overestimate productivity in such instances.

35 We present information about the balancing property in the Appendix. There, we also presentresults for the first-stage probit regressions.

36 We implement the matching algorithm using the Stata program psmatch2 written by Edwin Leuvenand Barbara Sianesi. Note that, even though we estimate the probit model pooled across all years, weensure exact matching by year in the second step of the matching approach.

17

where N refers to the number of treated firms, Ci to the set of control firms matched toeach treated firm i = 1, ..., N , and wij is a weight such that wij = 1

NCi

if j ∈ Ci and zero

otherwise, with NCi denoting the number of control firms. Below, we compute ATTs for

levels of the performance measures indicated by Hypothesis 3. We also consider alterna-tive variable transformations, and investigate the effects at various time horizons. To beprecise, besides level effects, we also consider responses of log changes. Importantly, com-puting an ATT from differences in outcome variables effectively combines propensity scorematching with difference-in-differences estimations. Indeed, Smith and Todd (2005) deemsuch a strategy especially appealing. While the approach still relies on the assumption of“selection on observables”, taking differences accounts for potential biases related to time-invariant firm characteristics. We thus also consider the change in outcomes computed asthe difference between post-treatment and pre-treatment log levels.

Moreover, we compute the treatment effects at various time horizons relative to thetreatment year. Assuming that the import event occurs in period p = 0, we generateATTs for the periods p = −2, . . . , 0, . . . , 3. Computing ATTs for pre-treatment periodsfunctions as a placebo test. In other words, prior to treatment, we would expect that firmsdiffer neither in terms of the control variables (as confirmed by the balancing tests) northe outcome variables. Moreover, analyzing the ATTs at various post-treatment periodsallows us to investigate the persistence of the effects. We note, though, that this typeof analysis implies a more selected sample, since both treated and untreated firms haveto be active for additional periods. As a result, the number of observations decreaseswith increasing p. Generally, independent of the time horizon considered, matching isalways performed at the time a firm starts importing, in line with De Loecker (2007), forexample.37

Finally, note that we exploit the double robustness property of the regression adjustedmachting estimator. To be precise, we compute ATTs by means of weighted regressions,using sampling weights obtained from the matching approach, while controlling for thecovariates included in the first stage probit regression. In this way, we can ensure that theestimator is consistent if either the propensity score equation or the regression equationis correctly specified (see for instance Imbens and Wooldridge, 2009). We cluster thestandard errors at the firm-level in these regressions.38

Results (H2 and H3)

Table 4 presents the baseline results referring to equation (19). Estimating the equationwithout firm fixed effects (columns i to iii), we find that importing retailers differ signif-icantly from non-importers across all performance measures suggested by Hypothesis 3.In particular, the sales, profits, and markups of directly importing firms exceed those oftheir non-importing counterparts on average by 20%, 27%, and 3%, respectively (columnsi to iii).39 Significant differences are also visible when adding firm fixed effects to the re-

37 Note that we focus on the first import event of firms in our sample and do not impose any restrictionson import activity in consecutive periods.

38 See also Goldbach et al. (2017) for a similar approach in an analysis of the effect of foreign investmenton domestic investment.

39 Computed e.g. as 19.7=(exp(0.18)-1)*100 since the dependent variables are log transformed. As aresult, we consider only firms with positive profits in these estimations (around 9% of firms report zeroor negative profits).

18

gressions (columns iv to vi); i.e. a switch in import status from zero to one appears to beassociated with better performance indicators. However, while these results are clearly inline with our theoretical model, we cannot distinguish between the effects related to moreproductive and therefore larger and more profitable firms self-selecting into importing, asimplied by Hypothesis 2, and those that are caused by import activity, as suggested byHypothesis 3.

As explained above, we investigate the self-selection hypothesis by running regressionsin accordance with equation (20). The estimation results in Table 5 show that firmsbeginning to import in year t are indeed already larger and more profitable than non-importing firms in period t − 1 (column i to iii). These differences amount to 9%, 5%and 2% for sales, profits, and markups, respectively. Hence, the results clearly supportthe self-selection mechanism proposed by Hypothesis 2. Moreover, the results presentedin Table 5 may also accord with additional performance gains after the import event hasoccurred since performance differences are even larger in period t (columns iv to vi).

We try to shed some light on the potentially causal effect of importing on retail firmperformance by applying a PSM approach, while keeping in mind that PSM, too, is basedon a number of assumptions. To the extent that these assumptions are plausibly fulfilledin our context, the results presented in Table 6 indeed suggest that importing significantlyimpacts sales, profits and, markups of retail firms. According to these estimates, a firmthat starts importing enjoys roughly 8% higher sales compared to non-importing firms inthe year of import initiation (panel a). This sales premium lasts during the consecutiveperiods and amounts to 11% after three years. Cumulating these effects over time forfirms present in all periods implies sales gains of close to 30% compared to firms sourcingdomestically. We also obtain a significant ATT when considering the change in (thelog of) sales as an outcome variable. In particular, on impact, the estimated effect ofimporting amounts to 9 percentage points higher sales growth for import starters. Thiseffect is quite persistent when considering longer time horizons. Three years after importinitiation, the growth in sales relative to pre-treatment exceeds that of non-importers by10 percentage points.

Moreover, we find a positive impact of import activity on firms’ profits which increaseby more than 6% in p = 0 and exceed those of non-importing firms by more than 11% threeyears after import initiation. Regarding growth rates, significant effects are estimatedwhich amount to 9 percentage points higher profit growth on impact and as much as18 percentage points higher profit growth when considering long differences in periodp = 3. Similarly, we find quite persistent effects on the markups of retailers that startto import. The results in panel c suggest that markups exceed those of non-importingfirms by between 2% and 3% in periods p = 0 and later. Moreover, markup growth is alsoaffected by around one to two percentage points, even though the statistical significancedecreases with longer time horizons.

Finally, we note that the Appendix contains information showing that the matchingapproach is successful in generating samples of treated and untreated firms that are wellbalanced in terms of covariates. Moreover, results in Table 7 indicate that, after applyingthe PSM approach, sales, profits, and markups do not differ significantly between thestudied treated and untreated firms prior to the import event. This table does not containplacebo results for changes in log sales in p− 1, since such a regression is redundant given

19

that we condition on pre-treatment sales growth in the selection model.40 Overall, we thusfind both evidence for a selection effect and indications that importing has contributedto the observed differences in the performance of importing and non-importing retailers,which is in line with Hypotheses 2 and 3.41

4.2 Import activity and consolidation into chains (H4)

Hypothesis 4 states that a reduction in the trade cost induces firms to add shops and thusbecome a retail chain.

Empirical approach (H4)

We investigate Hypothesis 4 by applying a similar methodology as before. We beginby estimating models as depicted by equation (19) for two different outcome variables,namely a dummy variable indicating chain retailers and a dummy variable taking a valueof one if a firm opens a new shop. Note that we estimate linear probability models inorder to be able to include firm fixed effects. These models indicate whether importingretailers have a higher probability of being a chain and of opening a shop. Moreover, thespecification with firm fixed effects relates changes in these probabilities to changes inimport status.

In addition, we apply a propensity score matching approach as outlined above in orderto gauge the direction of causality. Two things are worth noting here. First, in the caseof the chain dummy, we restrict the sample to firms that operated as single-shop firms int − 1, i.e. to firms for which the chain dummy equals zero before beginning to import.Hence we investigate whether import initiation changes the probability of becoming achain (in contrast to the probability of continuing as a chain). Second, when analyzingthe probability of opening a new shop, we add the (log of the) number of shops operated int− 1 to the first-stage probit regression. We are thus assessing whether import initiationaffects the probability of opening a new shop, comparing firms with a similar number ofshops prior to importing.

Results (H4)

The OLS results are presented in Table 8. As expected, we find that importing retailershave a higher probability of being a chain and of opening a new shop. The estimatedcoefficients imply probability differentials amounting to roughly 6 and 2 percentage points,respectively. The coefficients of interest shrink markedly when adding firm fixed effects,while they remain statistically and economically significant also in these regressions.

In Table 9, we present results from the matching approach in order to analyze thecausal relationship. On impact, we do not find a significant increase in the probability

40 Controlling for lagged sales growth in the probit regression (and consequently also when applying theregression adjusted matching estimator) also explains why we obtain the same coefficient when consideringthe log of sales in p− 1 and p− 2 as outcome variables.

41 In the Appendix, we present a series of robustness checks with respect to the matching approach.First, we add lagged (log-) levels of outcome variables to the first stage probit regressions. Second, weemploy exact matching by year and NACE four-digit sector. Third, instead of radius matching with atight caliper, we employ a Gaussian kernel in the matching step. Overall, our results are robust acrossthese checks.

20

of becoming a chain. However, when considering the results in consecutive periods, thereis some evidence of an effect two years after import initiation.42 The coefficient estimateimplies an effect of 1.6 percentage points, which is non-negligible when considering thatthe unconditional probability of being a chain amounts to only 10% in the overall sample.

Such a conclusion can also be drawn when focusing on the probability of opening anew shop. Indeed, import starters have a significantly higher probability of increasingthe number of shops in period p = 0, amounting to one percentage point. Moreover,we also observe that import starters have a significantly higher probability of opening anew shop two years after the import event has occurred. A coefficient estimate close to2 percentage points again implies a quantitatively meaningful effect, considering that, inthe data, the unconditional probability of adding a shop is only 3%. At the same time,we do not find significant differences in the probability of opening a new shop one yearbefore the import event has occurred.43 These results therefore suggest that importinghas indeed contributed to the consolidation of retailers into chains, which is in line withthe theoretical model’s prediction.

4.3 Import activity and firm exit (H1)

Hypothesis 1 derived from our theoretical model suggests that a trade cost reduction leadsto the exit of the least efficient firms.

Empirical approach (H1)

We use a different empirical approach than before to investigate this hypothesis. Thisis because, in this subsection, we do not only wish to analyze whether less productiveor non-importing retailers have a higher exit probability. Instead, we are interested inwhether these types of firms have a higher exit probability depending on trade integra-tion. To be precise, we test Hypothesis 1 by relating an indicator of firm-level exit (exitit)to industry-level imports. We are thus exploiting industry-level variation in import ac-tivity to assess firms’ exposure to trade integration. As noted before, this implies theassumption that increased trade integration is partly captured by observed changes inimports. We emphasize that the results presented in this subsection may not reflect acausal relationship, for instance, due to simultaneity issues.

We identify exits in the data by using a variable indicating the resignation date of afirm. The indicator variable exitit equals one if a firm has been active in year t − 1 andresigns from the market in year t. The comparison group of continuously active retailerscomprises firms that remain in the market during both years. We then estimate thefollowing model:

exitit = β0 + β1DIMPit + β2DSIZEit + β3LN(V IMPst) + β4DIMPit× (24)

LN(V IMPst) + β5DSIZEit × LN(V IMPst) + β6LN(SALESst) + αs + αt + εit,

42 Note that the chain (new shop) dummy is equal to one in periods p = 1 when a firm becomes achain (opens a new shop) in period p = 0 or p = 1. The dummies are coded equivalently for period p = 2and p = 3.

43 A placebo regression for the chain dummy is redundant since we restrict the analysis to comparingsingle-shop firms in the pre-treatment period.

21

where DIMPit is a dummy variable indicating a firm’s import status; DSIZEit is acategorical variable indicating whether firm i is among the smallest, i.e. least effi-cient, firms in a given year; LN(V IMPst) denotes the value of non-EU15 consumergoods imports of retail firms belonging to sector s in year t; DIMPit × LN(V IMPst)and DSIZEit × LN(V IMPst) are interaction terms; and LN(SALESst) measures totalindustry-level sales to control for other aggregate developments. Hence equation (24) al-lows us to investigate whether industry-level imports have a differential effect on the exitprobability of importing and non-importing firms or of small and large retailers.

We compute DSIZEit based on the size distribution (total assets) of firms in theestimation sample. Specifically, we define firms that belong to 25th percentile of thesize distribution as small. Note that we proxy for efficiency using a firm’s size ratherthan productivity because several firms do not report any employees, thus preventing thecomputation of labor productivity. Since these non-reporting firms tend to be small andexit the market rather frequently, we would lose relevant information if we relied on laborproductivity.44

Finally, note that the industry level s here refers to the 24 industries described insection 3 and in Table A.3 in the Appendix. Since our variable of interest (ln(V IMPst))varies over this dimension, we cluster the standard errors at the sector level (Moulton,1990). Moreover, we also present p-values derived from a wild bootstrap procedure (inbrackets), given that the number of clusters is relatively small (Cameron et al., 2008).45

Results (H1)

The estimation results in Table 10 first of all show that importing firms have a lowerand small firms have a higher probability of exiting the market.46 The coefficient of theimport dummy suggests a 2.6 percentage point lower exit probability and that of the sizedummy a 16.7 percentage point higher exit probability for the respective types of firms.These results thus fit our theoretical model, in which exit occurs at the lower end of theretailer size distribution and thus among retailers that do not import.

The results in column (i) suggest that an increase in industry-level imports is, in gen-eral, associated with an increase in the exit probability of non-importing firms (coefficientβ3). However, while the interaction term with a firm’s import status has the expectednegative coefficient, it is not statistically significant. On the other hand, the results incolumn (ii) are more in line with Hypothesis 1. In this case, the coefficient for industry-level imports does not suggest a significant relationship with the exit probability of largerretailers. At the same time, the interaction between the size dummy and industry-levelimports is positive and significant, which suggests that an increase in aggregate importsis associated with a higher exit probability of small retailers. This finding is confirmedby the results shown in column (iii), where both interaction terms are included simulta-neously. Note that the implied effect is relatively small. Adding together the coefficients

44 Numerous empirical studies document that firm size is highly correlated with productivity. Moreover,our theoretical model suggests a direct link between size and productivity. Furthermore, as mentionedin footnote 34, a measure of labor productivity may be problematic, especially for smaller retailers, sincethe variable ignores the work done by firm owners.

45 We implement this procedure in Stata using the cgmwildboot routine written by Judson Caskey.The null hypothesis is always imposed.

46 Note that the import variable LN(V IMPst) is mean-centered.

22

of LN(V IMPst) and the interaction term, i.e. computing β3 + β5, the estimated effectsamounts to less than 0.02 percentage points in response to an increase in sector-level im-ports of 1%. Overall, we thus conclude that our results are broadly in line with Hypothesis1.

4.4 Import activity and aggregate productivity and markups(H5)

Our model indicates that the adjustments in firm performance induced by direct importactivity have consequences at the aggregate retail market level. According to Hypothesis5, we would expect to find an increase in aggregate productivity, while the effect on theaggregate markup is not clear-cut and depends, inter alia, on the productivity distributionof the firms.47

Empirical approach (H5)

We take these predictions to the data by conducting a set of regressions at the local retailmarket level, where a local market is defined by region r and sector s. As explainedabove, we observe certain variables such as sales, total assets, productivity or markupsonly at the firm level. Hence, for a chain retailer we have to distribute these variablesacross its locations in different regions. For sales and assets, we do this on the basis ofthe employment share of each shop in a chain’s total employment. For productivity andmarkups, we assume that they are similar across shops belonging to the same firm. In asecond step, we aggregate across shops present in a local retail market to obtain variablesvarying by local retail market and time. For productivity and markups, we compute aweighted average using a shop’s sales as weight.48

We then relate market level characteristics, namely productivity and markups, toindustry-level consumer goods imports:

yrst = β0 + β1LN(V IMPst) + β2(xrs,t−1) + αrs + αt + εit, (25)

where yrst is a market-level outcome (log-transformed); LN(V IMPst) refers to the valueof industry-level imports (as before); and xrs,t−1 are lagged covariates to control for localretail market characteristics, namely size (measured as the sum of assets of all shops ina market) and the average wage (measured as total wage bill over total employment ina market). By including local market fixed effects (αrs) in these regressions, we exploitwithin market variation to estimate the relationship between imports and the outcomeof interest. As before, we rely on OLS estimation so that the following results should be

47 To put this hypothesis and the results presented below for retailers into perspective, it is useful tocompare them to studies on the effects of trade liberalization on average markups and productivity inmanufacturing. De Loecker et al. (2016) show, using Indian data, how a decrease in input tariffs inducesmanufacturers to raise their markups and thus pass only part of the tariff reduction on to consumers.This corresponds to the direct effect on markups explained in our Eq. (16). A similar effect is shown byBrandt et al. (2017) using Chinese data.

48 Notice a slight discrepancy here between our empirical approach and the theoretical model. In amonopolistically competitive model, a firm, or, in our case, a shop, is implicitly assumed to have negligibleweight in the industry.

23

interpreted as representing suggestive evidence for the presence of certain relationships inthe data without drawing any causal conclusions. Note that we present results from bothunweighted and weighted regressions, where weights are based on regional sales shares ofeach industry. Further note that, as before, we cluster the standard errors at the industrylevel and also report p-values derived from a wild bootstrap procedure in brackets toaccount for the relatively small number of clusters.

Results (H5)

Table 11 contains the estimation results, where the first two columns refer to unweightedregressions. The results suggest that industry-level imports are indeed positively relatedto market-level productivity (column i). Moreover, the coefficient for aggregate markupsdisplays a negative relationship (column ii). While the analytical standard errors implythat these coefficients are (weakly) statistically significant, p-values derived from the wildbootstrap procedure are considerably larger, indicating that there is great uncertaintyaround these estimates.

When focusing on the weighted regressions, we no longer find a negative coefficientfor markups (column iv). Instead, the coefficient is positive, though close to zero andestimated very imprecisely. Overall, these results therefore fit our model, which alsopredicts that aggregate markup effects are ambiguous. In contrast, the coefficient forproductivity remains positive also when using weighted OLS (column iii). While this isalso expected from the model, we note that the coefficient’s p-value increases to 0.13 inthe case of the wild bootstrap approach.

4.5 Import activity and local-market-level concentration (H6and H7)

Finally, our model provides conditional predictions about the components of a local mar-ket’s Herfindahl index, i.e. the number of shops, as well as the mean and the standarddeviation of local retail market sales. For small enough fixed costs of importing, a re-duction in trade costs would lower the number of shops (H6) and decrease the coefficientof variation of retail sales (H7). Fewer shops, ceteris paribus, imply a greater Herfindahlindex of market concentration, but a lower coefficient of variation of retail sales leads toa smaller index. Hence, in theory, changes in trade costs have an ambiguous effect on theHerfindahl index.

Empirical approach (H6 and H7)

We investigate these prediction by resorting to equation (25), where yrst now refers to theHerfindahl index or one of its components.49

49 Note that some retail firms may operate several shops in a given local market. But shops belonging tothe same firm probably do not compete with each other in prices. A similar assumption is made by Holmes(2011) and supported by evidence from Holmes on considerable self-cannibalization among Walmartstores. Treating shops by the same firm as independent shops would thus lead to an underestimation oflocal market concentration. To obtain a more accurate measure of market concentration, we thereforetreat the shops of a given firm in region r and industry s as a single shop. The number of shops in alocal market is therefore equal to the number of firms operating a shop or shops in this market.

24

Results (H6 and H7)

Table 12 contains the estimation results. In line with Hypothesis 6, we find that anincrease in market-level imports is negatively related to the number of retailers presentin a market (column i). On the other hand, the relationship between industry-levelimports and average market-level sales (column ii) as well as the relationship betweenindustry-level imports and the standard deviation of market-level sales (column iii) areestimated very imprecisely.50 Finally, we find that a rise in imports is associated with anincrease in the Herfindahl index at the local retail market level and thus a rise in localmarket concentration (column iv). The results are broadly similar when using a weightedregression approach (columns v to viii). By our estimates, a 1% increase in industry-leveldirect imports raises the Herfindahl index of local retail market concentration, ceterisparibus, by up to 0.042%. This is not a small effect, especially when considering thataverage market-level direct imports in the estimation sample more than doubled duringthe sample period and that, as shown in Table 1, the Herfindahl index is at a high levelinitially.

4.6 The role of indirect imports

The bottom line of our empirical analysis so far is that we tend to find support for ourtheoretical model. This is especially true when estimations are conducted at the firmlevel. But, even in the case of market-level regressions, we find relationships that are inline with model predictions, especially in cases where the theoretical model yields unam-biguous comparative statics. The empirical results hence provide considerable evidencethat observed changes in the performance of retailers and in structural changes in retailindustries are driven at least partly by the increase in retailers’ direct importing activities.

But how sure can we be that our findings are only related to retailers’ direct imports,as suggested by our theoretical model? Could we not find similar effects at least for retailmarket structure from indirect imports of consumer goods, i.e. imports by wholesalersand other firms outside the retail sector? As far as our theoretical model is concerned, itwould be straightforward, if somewhat tedious, to introduce indirect imports as an addi-tional option for firms; this could be done without fundamentally changing the economicmechanism driving the results, at least as long as the wholesale sector is not as efficientin distributing imported goods as direct importers (see, for instance, Raff and Schmitt(2012) for further details). The bigger problem is that, as noted before, we do not ob-serve firms’ domestic sourcing activities and thus do not know whether they import goodsindirectly, for instance, through wholesalers.

However, our data allow us to compute a proxy for indirect imports at the industrylevel. Our strategy for generating such a proxy for indirect imports is based on exploitingtwo types of information available in the trade data. First, we observe a retail industry’smix of directly imported consumer good varieties. We define a retail industry’s direct

50 Notice, however, that the theoretical model itself does not deliver clear-cut predictions regardingthe effect of imports on the number of shops and the first two moments of local retail sales. Specifically,Hypotheses 6 and 7 only hold for sufficiently small fixed costs of importing.

25

import share of a particular variety as:

shscjt =V IMPscjtV IMPst

,

where V IMPscjt denotes direct imports by retail industry s of variety cj (defined bya country c and product j combination) in year t.51 Second, we know total imports ofeach consumer goods variety conducted by non-retail firms, which we denote as V IMPO

cjt,where superscript O indicates that these imports are generated by firms outside of theretail industry. Our proxy for a retail industry’s indirect imports is then computed as:

V IMPOst =

∑cj

shscjtV IMPOcjt.

An important assumption underlying this measure is that an industry’s direct imports areinformative for its indirect imports. In particular, this variable does not capture indirectimports of varieties that are not imported by any directly importing retailer. Hence, thisvariable indeed is a rather rough proxy and the following results can only provide somesuggestive evidence.

Table 13 presents the exit regressions, while Tables 14 and 15 present the results formarket-level regressions. In each case, we run similar regressions as before, while we nowalso add the proxy for indirect imports to the models (and the corresponding interactionsin case of the exit regressions). Interestingly, with one exception (column vi in Table15), the variable V IMPO

st (or an interaction based on this variable) does not exhibit astatistically significant effect on an outcome of interest. Moreover, including this variablein the regressions has hardly any impact on the coefficient(s) related to direct imports.Thus, these results do indeed suggest that our findings are driven by retailers’ directimport activities and are not merely related to more consumer goods imports coming toDenmark in general.

5 Conclusions

The paper used Danish microdata for the period 1999 to 2008 to examine the impact ofthe rapid growth of consumer goods imports on retail market performance and structure.Our results suggest that larger and more profitable retailers do not only self-select intoimporting, but that retailers’ import activities are responsible for some of the performancedifferences observed between importing and non-importing firms. In particular, based ona propensity score matching approach, we find that retailers that begin to import mayhave 8% greater sales, 6% greater profit, and 2% greater markups after beginning toimport compared to non-importing firms. These differences are quite persistent and alsohold for the growth rates of the considered performance measures. For instance, we findthat cumulative sales of import starters may be up to 30% higher on average after threeyears than for comparable non-importers.

Regarding the other stylized facts we emphasized in the introduction, in particular the

51 The product dimension refers to the HS six-digit level. We concord these product codes over timefollowing Van Beveren et al. (2012).

26

consolidation of retailers into chains, the exit of small retailers, and greater retail marketconcentration, we offer several pieces of evidence that link these processes to greaterdirect imports of consumer goods. First, in our firm-level estimations we find evidencesuggesting that import initiation raises the probability of a retailer becoming a chain andadding new shops to its portfolio. We also detect indications for a positive relationshipbetween the probability of exiting the market for small retailers and increases in sector-level direct imports. Moreover, we present evidence for a positive conditional correlationbetween greater direct imports and higher local retail market concentration as measuredby the Herfindahl index. This increase in market concentration appears to be driven bya reduction in the number of retail firms associated with greater direct imports.

Taken together, these empirical results provide support for the economic mechanismexplored in our theoretical model, namely that performance differences between importersand non-importers as well as structural changes in retailing are, inter alia, driven bydirect importing: Due to economies of scale associated with direct importing, only bigretailers and retail chains can afford direct imports, which means that they benefit morefrom trade cost reductions than small retailers that do not have access to these imports.We obtain additional evidence in favor of this mechanism by examining how indirectimports of consumer goods, i.e. imports by wholesalers and other firms outside the retailindustry, affect retail markets in Denmark. Almost without exception, our proxy forindirect imports does not exhibit a statistically significant effect on firm exit, on aggregateretail productivity and markups, on the number of retailers, or on the Herfindahl indexof retail market concentration.

We conclude that, viewed from a positive perspective, our model provides a plausiblemechanism to explain some of the changes in retailing observed in Denmark and otheradvanced economies. Our paper thus complements the existing literature on retailing thatgenerally attributes observed changes in retailing to technological change.

What, from a normative perspective, do our results imply for social welfare? In ourtheoretical model, social welfare, W , is approximately equal to:

W ≈ 2Rn

∫ 1/2n

0

(v − τx)Ldx (26)

−Rn(∫ cI

0

[(t+ c)qI(c)

]dG(c) +

∫ cD

cI

[(w + c)qD(c)

]dG(c)

)(27)

−RnE[G(cI)FI + FE] + (R− 1)nEG(cs)[FI + FE − FS], (28)

where the number of active shops in a given regional market, n, is related to the number ofentrants, nE, through n = G(cD)nE, and v is a consumer’s reservation price for one unitof the aggregate consumption good. Thus welfare is approximately equal to consumersurplus net of the travel cost paid by consumers to reach the nearest retailer (26), minusthe expected (variable) trade and production costs (27), minus the sunk cost of entry andthe fixed costs of importing and operating shops in additional markets (28).52

52 The expression is an upper bound on welfare, since the first term assumes that consumers on averagetravel 1/4 of the “distance” between any two shops, which would be the case if all retailers charged thesame price. However, since prices vary across retailers in our setting, travel costs are strictly greater thanshown here.

27

Our empirical results indicate that increased consumer goods imports are negativelyrelated to the number of retail firms, which may thus imply an increase in travel distancesfor consumers to reach a shop. As explained by Lagakos (2016), the social costs of greaterdistance are probably small in an advanced country like Denmark, where households tendto own cars and have access to efficient public transport. A greater welfare impact foran advanced economy can be expected to come from lower import prices, as retailersturn to imports instead of domestically sourced goods, and from the trade-induced shiftof market share towards the bigger retailers and retail chains that import directly. Thesourcing of lower-priced goods from abroad is, of course, a classic gain from trade. Theincrease in industry productivity stemming from a market share reallocation towards thebig and therefore typically more efficient retailers is reminiscent of the effects highlightedin heterogeneous firm models of trade, such as Melitz and Ottaviano (2008). But, in thecurrent paper, this productivity gain arises in retailing not in manufacturing, and thusrepresents an additional gain from trade that is not captured by traditional trade models.In the data we indeed tend to find some evidence for such a gain, namely a positiverelationship between labor productivity and industry-level imports, which is, however,estimated somewhat imprecisely.

Another non-traditional gain from trade highlighted by our model consists of thepotential cost savings associated with the trade-induced consolidation of shops into chains.These cost savings arise because chains can spread the fixed costs of importing (FI) andthe overhead costs that arise at the firm rather than the shop level (FE−FS) across theirshops. A quantification of these welfare gains is beyond the current paper, but the strongincrease in the number of retail chains and the fact that these chains are responsible forthe lion’s share of import activity in Danish retailing are indicative of substantial gains.That these gains are likely to be big is also suggested by Holmes (2011), who estimatesthe savings on distribution costs that Walmart realized when expanding the number ofits stores, and by Holmes and Singer (2017), who find large savings in import costs forbig retail chains like Walmart.

Any future study of the welfare gains from trade-induced changes in retailing wouldhave to take a position on several additional aspects that we have not touched upon in ouranalysis. First, in many countries, retailing is highly regulated. Regulation, for instanceof shop size or location, may limit the extent to to which the large, productive retailers inparticular can adjust to trade (Eckel, 2009, and Raff and Schmitt, 2012). In the case ofDenmark, retailers are subject to a zoning regulation that limits the maximum shop size.53

The OECD indicator on the regulation of large outlets suggests that the zoning regulationin Denmark is not among the most restrictive and had been loosened somewhat at theend of our sample period.54 But there are other countries, such as Australia, Canada andthe United States, that do not systematically impose such regulations and therefore mayhave experienced bigger retail market adjustments to trade than Denmark. That this

53 From 1997, the zoning law prohibited grocery stores larger than 3,000 sq.m (since 2007 larger than3,500 sq.m) in urban areas; in rural areas, the size limit is 1,000 sq.m. For non-grocery stores, the zoningregulation implies a size limit of 2,000 sq.m, while large cities can grant three exceptions to this rule peryear.

54 The indicator is based on the following survey question: “If compliance with regulation especiallydesigned for large outlets is required for any type of product, what is the threshold surface limit for theselaws or regulations to apply (in m2)?” See Koske et al. (2015) for a comparison of this indicator acrossOECD countries.

28

may indeed be the case is indicated by Basker and Van (2010b) and Holmes and Singer(2017) who document how Walmart has come to dominate both retailing and consumergoods trade in the United States.

Second, we did not examine changes in the product assortment of retailers, except tonote that, over the sample period, Danish retailers substantially expanded the numberof product varieties that they import directly. In the case of consumer goods, retailers,of course, play an essential role in determining the number of product varieties thatconsumers get to choose from, and their decisions regarding the product range are thuslikely to have important welfare effects. Several recent theoretical studies explain howretailers expand their product range in response to trade liberalization, but they also showthat the welfare effects may not be straightforward. Eckel (2009) explores the possibilitythat expanding the product range may lead to higher retail prices, thus potentially hurtingconsumers. He also shows that retailers may offer too little product variety compared tothe social optimum. Raff and Schmitt (2016b) argue that the product range offered inequilibrium may be too broad. Their model, however, indicates that a consolidation ofshops into retail chains would bring product variety closer to the socially optimal level,thus leading to additional gains that are directly related to the increase in the number ofchains.

Third, we did not explore the relationship between retailers and manufacturers. Asretailers are getting bigger and, as Feenstra and Hamilton (2006) suggest, gain powerespecially vis-a-vis overseas manufacturers, they may be tempted to use this power toextract greater rents from manufacturers. Raff and Schmitt (2009) examine this issueand show how the exercise of buyer power by retailers may lead to smaller gains fromtrade than indicated by models in which market power rests with manufacturers.

References

[1] Ackerberg, D. A., Caves, K., Frazer, G., 2015. Identification properties of recentproduction function estimators. Econometrica 83, 2411–2451.

[2] Antoniades, A., Zaniboni, N., 2016. Exchange rate pass-through into retail prices.International Economic Review 57, 1425–1447.

[3] Atkin, D., Faber, B., Gonzalez-Navarro, M., 2018. Retail globalization and householdwelfare: Evidence from Mexico. Journal of Political Economy 126, 1-73.

[4] Basker, E., Van, P. H., 2010a. Putting a smiley face on the dragon: Wal-Mart as cat-alyst to US-China trade. University of Missouri, Department of Economics WorkingPaper 07-10.

[5] Basker, E., Van, P. H., 2010b. Imports “r” us: Retail chains as platforms fordeveloping-country imports. American Economic Review 100, 414–418.

[6] Basker, E., Klimek, S., Van, P. H., 2012. Supersize it: The growth of retail chainsand the rise of the ‘big-box’ store. Journal of Economics and Management Strategy21, 541–582.

29

[7] Bernard, A. B., Jensen, J. B., Redding, S., Schott, P., 2010a. Wholesalers and retail-ers in US trade. American Economic Review 100, 408–413.

[8] Bernard, A. B., Jensen, J. B., Redding, S., Schott, P., 2010b. Wholesalers and re-tailers in US Trade (Long Version). NBER Working Paper 15660.

[9] Brandt, L., Van Biesebrock, J., Wang, L., Zhang, Y., 2017. WTO accession andperformance of Chinese manufacturing firms. American Economic Review 107, 2784–2820.

[10] Caliendo, M., Kopeinig, S., 2008. Some practical guidance for the implementation ofpropensity score matching. Journal of Economic Surveys 22, 31–72.

[11] Campa, J. M., Goldberg, L. S., 2006. Distribution margins, imported inputs, and thesensitivity of the CPI to exchange rates. NBER Working Paper 12121.

[12] Cameron, A. C., Gelbach, J. B., Miller, D. L., 2008 Bootstrap-based improvementsfor inference with clustered errors. The Review of Economics and Statistics 90, 414–427.

[13] De Loecker, J., Goldberg, P.K., Khandelwal, A.K., Pavcnik, N., 2016. Prices,markups, and trade reform. Econometrica 82, 445-510.

[14] De Loecker, J., 2007. Do exports generate higher productivity? Evidence from Slove-nia. Journal of International Economics 73, 69–98.

[15] De Loecker, J., Warzynski, F., 2012. Markups and firm-level export status. AmericanEconomic Review 102, 2437–2471.

[16] Eckel, C., 2009. International Trade and Retailing. CESifo Working Paper 2597.

[17] Feenstra, R. Hamilton, G., 2006. Emergent economies, divergent paths: Economicorganization and international trade in South Korea and Taiwan. Cambridge Uni-versity Press, Cambridge (UK).

[18] Foged, M., Peri G., 2015. Immigrants’ effect on native workers: New analysis onlongitudinal data. American Economic Journal: Applied Economics 8, 1–34.

[19] Foster, L. S., Haltiwanger, J., Klimek, S., Krizan, C.J., Ohlmacher, S., 2016. Theevolution of national retail chains: How we got here. in: Emek Basker (ed.), Hand-book on the Economics of Retailing and Distribution. Edward Elgar Publishers,Cheltenham, 7–37.

[20] Gereffi, G., Frederick, S., 2010. The global apparel value chain, trade and the crisis:Challenges and opportunities for developing countries. World Bank Policy ResearchWorking Paper 5281.

[21] Goldbach, S., Nagengast, A. J., Steinmuller, E., Wamser, G., 2017. The effect ofinvesting abroad on investment at home: On the role of technology, tax savings, andinternal capital markets. Deutsche Bundesbank Discussion Paper 14/2017.

30

[22] Hall, R. E., 1988. The relation between price and marginal cost in U.S. industry.Journal of Political Economy 96, 921–947.

[23] Holmes, T. J., 2001. Bar codes lead to frequent deliveries and superstores. RandJournal of Economics 32, 708–725.

[24] Holmes, T. J., 2011. The diffusion of Wal-Mart and economies of density. Economet-rica 79, 253–302.

[25] Holmes, T. J., Singer, E. 2017. Indivisibilities in distribution. Federal Reserve Bankof Minneapolis, Working Paper 739.

[26] Hummels, D., Jørgensen, R., Munch, J., Xiang, C., 2014. The wage effects of off-shoring: Evidence from Danish matched worker-firm data. American Economic Re-view 104, 1597–1629.

[27] Imbens, G. W., Wooldridge, J. M., 2009. Recent developments in the econometricsof program evaluation. Journal of Economic Literature 47, 5–86.

[28] Innes, R., 2006. Entry deterrence by non-horizontal merger. Journal of IndustrialEconomics 54, 369–395.

[29] Jarmin, R. S.,Klimek, S. D., Miranda, J. 2009. The role of retail chains: National,regional and industry results. in: Dunne, T., Jensen, J., Roberts, M. (Eds.), ProducerDynamics: New Evidence from Micro Data, NBER Books, Cambridge (US), pp. 237–262.

[30] Koske, I., Wanner, I., Bitetti, R., Barbiero, O., 2015. The 2013 update of the OECD’sdatabase on product market regulation–Policy insights for OECD and non-OECDcountries. OECD Economics Department Working Paper 1200.

[31] Lagakos, D., 2016. Explaining cross-country productivity differences in retail trade.Journal of Political Economy 124, 579–620.

[32] Levinsohn, J., Petrin, A., 2003. Estimating production functions using inputs tocontrol for unobservables. Review of Economic Studies 70, pages 317–341.

[33] Maican, F., Orth, M., 2017. Productivity dynamics and the role of ‘big-box’ entrantsin retailing. Journal of Industrial Economics 65, 397–438.

[34] Melitz, M. J., I.P. Ottaviano, G. I. P., 2008. Market size, trade and productivity.Review of Economic Studies 75, 295–316.

[35] Moulton, B. R., 1990. An illustration of a pitfall in estimating the effects of aggregatevariables on micro units. Review of Economics and Statistics 72, 334–338.

[36] Nordas, H. K., 2008. Gatekeepers to consumer markets: The role of retailers in inter-national trade. International Review of Retail, Distribution and Consumer Research18, 449–472.

31

[37] Raff, H., Schmitt, N., 2009. Buyer power in international markets. Journal of Inter-national Economics 79, 222–229.

[38] Raff, H., Schmitt, N., 2012. Imports and the structure of retail markets. CanadianJournal of Economics 45, 1431–1455.

[39] Raff, H., Schmitt, N., 2016a. Retailing and international trade. in: Basker, E. (Ed.),Handbook on the Economics of Retailing and Distribution. Edwar Elgar Publishers,Cheltenham, pp. 157–179.

[40] Raff, H., Schmitt, N. 2016b,. Manufacturers and retailers in the global economy.Canadian Journal of Economics 49, 685–706.

[41] Rosenbaum, P. R., Rubin, D. B., 1983. The central role of the propensity score inobservational studies for causal effects. Biometrika 70, 41–55.

[42] Rosenbaum, P. R., Rubin, D. B., 1985. Constructing a control group using multivari-ate matched sampling methods that incorporate the propensity score. The AmericanStatistician 39, 33–38.

[43] Smeets, V., Warzynski, F., 2013. Estimating productivity with multi-product firms,pricing heterogeneity and the role of international trade, Journal of InternationalEconomics 90, 237–244.

[44] Smith, A. J., E. Todd, P. E., 2005. Does matching overcome LaLonde’s critique ofnonexperimental estimators? Journal of Econometrics 125, 305–353.

[45] Triplett, J., B. Bosworth, B., 2004. Productivity in the US services sector: Newsources of economic growth. Brookings Institution Press, Washington D.C..

[46] Van Beveren, I., Bernard, A.B., Vandenbussche, H., 2012. Concording EU trade andproduction data over time. NBER Working Papers 18604.

[47] Wagner, J., 2007. Exports and productivity: A survey of the evidence from firm-leveldata, World Economy 30, 60–82.

[48] Waterson, M., 1984. Economic theory of industry. Cambridge University Press, Cam-bridge (UK).

[49] Zentes, J., Hilt, C., Domma, P., 2007 Handelsmonitor Spezial: Global Sourcing imEinzelhandel. Deutscher Fachverlag, Frankfurt a.M..

32

Table 1: Retail firms over time

i ii iii iv v vi vii

Share of Share ofretail imports Average

firms in by retail imports Share oftotal im- firms Share of of im- im- Herfin-porting in total Average imports porting porting dahlfirms imports sales in sales firms firms index

1999 12.2 9.2 10.0 3.5 5.0 7.1 0.242000 11.8 9.2 10.1 4.0 5.6 7.2 0.242001 11.9 8.9 10.3 4.0 5.8 7.0 0.252002 11.6 7.5 10.6 3.4 4.2 8.6 0.252003 12.1 8.6 10.3 4.1 4.5 9.5 0.252004 12.4 9.6 10.6 4.4 4.8 9.8 0.252005 12.6 10.9 10.7 5.4 5.7 10.1 0.242006 12.4 13.5 11.0 7.0 7.5 10.3 0.242007 12.5 13.7 11.5 7.7 8.2 10.8 0.242008 12.1 13.8 11.5 7.7 8.3 10.8 0.24

Notes: Sales and import values in million DKK in prices from 2000. The Herfindahl index is computed as the weightedaverage across local retail markets (weights are a market’s yearly sales share). Imports refer to consumer goods imports.

33

Table 2: Retail chains over time

i ii iii iv

Share in Share inNo. of retail Share of retailchains sales importers imports

1999 709 55.2 26.9 87.72000 755 56.9 24.8 86.92001 793 57.7 23.0 87.42002 853 63.6 26.0 86.72003 876 62.2 26.0 88.32004 960 61.7 26.7 88.42005 1027 64.1 27.0 89.42006 1094 63.5 26.6 91.02007 1046 63.8 28.0 91.32008 1264 64.1 23.4 92.3

Notes: Table is based on firms present in the data sets FIRM and IDAS. A chain is defined as a firm with at least twoshops. Imports refer to consumer goods imports.

Table 3: Direct imports by chain and single-shop retailers (in 2005)

i ii iii iv

Mean Meandomestic em- Mean No. of

sales ployees imports firms

Chain - importer 337.6 173.1 33.7 268Chain - not importer 38.5 23.4 0.0 740Not chain - importer 8.2 4.9 1.1 944Not chain - not importer 6.6 4.0 0.0 8560

Notes: Data refer to retail firms with at least one employee that were active in 2005 and present in the data sets FIRMand IDAS. Columns i and iii in million DKK. Imports refer to consumer goods imports.

34

Table 4: Direct importing and firm performance - OLS (H2, H3)

i ii iii iv v vi

Dependent variable: Log of Sales Profits Markup Sales Profits Markup

DIMPit 0.180 0.241 0.033 0.110 0.108 0.011(0.000) (0.000) (0.000) (0.000) (0.000) (0.015)

LN(EMPLOY EESit−1) 0.847 0.557 -0.003 0.334 0.188 0.000(0.000) (0.000) (0.165) (0.000) (0.000) (0.990)

Observations 92,021 79,251 75,370 92,021 79,251 75,370R-squared 0.749 0.295 0.079 0.881 0.624 0.548

Firm fixed effects no no no yes yes yes

Mean of dependent variable 15.325 12.600 0.150 15.325 12.600 0.150

Notes: Firm-level regressions containing sector-year dummies. P-values referring to standard errors clustered at the firmlevel in parentheses. DIMPit is a dummy variable taking on unity if a firm imports consumer goods from outside EU15.LN(EMPLOY EESit−1) refers to number of employees.

Table 5: Import starters and firm performance - OLS (H2)

i ii iii iv v vi

Dependent variable: Log ofLagged (t-1) Contemporaneous (t)

sales profits markup sales profits markup

STARTit 0.085 0.051 0.020 0.154 0.108 0.029(0.000) (0.074) (0.006) (0.000) (0.000) (0.000)

LN(EMPLOY EESit−1) 0.838 0.515 -0.008 0.855 0.536 -0.012(0.000) (0.000) (0.000) (0.000) (0.000) (0.000)

Observations 78,220 57,556 60,752 78,220 57,556 60,752R-squared 0.827 0.288 0.211 0.738 0.282 0.084

Mean of dependent variable 15.305 12.633 0.192 15.238 12.590 0.147

Notes: Firm-level regressions containing sector-year dummies. P-values referring to standard errors clustered at the firmlevel in parentheses. STARTit is a dummy variable taking on unity if a firm did not import in periods t− 2 and t− 1 andbegins to do so in period t; it is zero for firms that have not imported during all three periods. Imports refer to consumergoods from outside EU15. LN(EMPLOY EESit−1) refers to number of employees.

35

Table 6: Direct importing and firm performance - matching (H3)

p p+1 p+2 p+3

Panel a: Domestic sales

LN(SALES)0.084 0.084 0.104 0.110

(0.000) (0.000) (0.000) (0.000)

∆LN(SALES)0.089 0.093 0.102 0.103

(0.000) (0.000) (0.000) (0.000)Number of treated 1,355 1,152 943 766

Observations 69,025 54,910 41,469 31,256

Panel b: Profits

LN(PROFITS)0.064 0.097 0.090 0.115

(0.015) (0.002) (0.012) (0.004)

∆LN(PROFITS)0.090 0.122 0.115 0.180

(0.001) (0.000) (0.003) (0.000)Number of treated 1,041 828 645 488

Observations 50,329 36,954 26,765 18,496

Panel c: Markups

LN(MARKUPS)0.019 0.028 0.026 0.030

(0.006) (0.000) (0.004) (0.002)

∆LN(MARKUPS)0.012 0.018 0.017 0.017

(0.067) (0.020) (0.079) (0.100)

Number of treated 1,073 872 677 523Observations 52,598 38,976 27,854 19,863

Notes: P-values referring to standard errors clustered at the firm level in parentheses. Coefficients refer to a dummyvariable indicating import initiation from non-EU15 markets (STARTit) obtained from weighted firm-level regressions withweights derived by propensity score matching. Regressions control for covariates included in first-stage probit regression.∆ indicates that changes in the outcome variable are considered. Note that these differences are computed relative topre-treatment status (i.e. relative to p-1). p refers to the year of treatment. Table A.11 in the Appendix presents means ofthe outcome variables considered in this table.

36

Table 7: Direct Importing and firm Performance - matching (H3) - placebo

p-2 p-1

Panel a: Domestic Sales

LN(SALES)-0.005 -0.005(0.564) (0.564)

Number of treated 1,355 1,355Observations 69,025 69,025

Panel b: Profits

LN(PROFITS)0.006 -0.025

(0.815) (0.191)Number of treated 911 1,041Observations 42,536 50,329

∆LN(PROFITS)-0.031(0.266)

Number of treated 911Observations 42,536

Panel c: Markups

LN(MARKUPS)-0.001 0.007(0.943) (0.231)

Number of treated 912 1,073Observations 43,397 52,598

∆LN(MARKUPS)0.008

(0.260)Number of treated 912Observations 43,397

Notes: P-values referring to standard errors clustered at the firm level in parentheses. Coefficients refer to a dummyvariable indicating import initiation from non-EU15 markets (STARTit) obtained from weighted firm-level regressions withweights derived by propensity score matching. Regressions control for covariates included in first-stage probit regression.∆ indicates that changes in the outcome variable are considered. Note that these differences are computed using the p− 2samples, while considering the difference in the log of outcomes in p− 1 and p− 2. p refers to the year of treatment.

37

Table 8: Direct importing and consolidation into chains - OLS (H4)

i ii iii iv

Dependent variableChain New-shop Chain New-shop

dummy dummy dummy dummy

DIMPit 0.058 0.024 0.011 0.007(0.000) (0.000) (0.023) (0.075)

LN(EMPLOY EESit−1) 0.138 0.036 0.088 0.004(0.000) (0.000) (0.000) (0.068)

Observations 88,319 87,437 88,319 87,437R-squared 0.252 0.062 0.750 0.173

Firm fixed effects no no yes yes

Mean of dependent variable 0.098 0.025 0.098 0.025

Notes: Firm-level regressions containing sector-year dummies. P-values referring to standard errors clustered at the firmlevel in parentheses. DIMPit is a dummy variable taking on unity if a firm imports consumer goods from outside EU15.LN(EMPLOY EESit−1) refers to number of employees.

Table 9: Direct importing and consolidation into chains - matching (H4)

p-1 p p+1 p+2 p+3

Chains

CHAIN DUMMY0.002 0.004 0.016 0.011

(0.586) (0.486) (0.068) (0.303)

Number of treated 1,074 887 723 582Observations 58,318 44,968 33,323 24,461

New shops

NEW-SHOP DUMMY-0.002 0.010 0.010 0.022 0.019(0.625) (0.054) (0.154) (0.017) (0.085)

Number of treated 1,285 1,319 1,091 894 722Observations 64,474 66,154 51,284 38,442 28,877

Notes: P-values referring to standard errors clustered at the firm level in parentheses. Coefficients refer to a dummy variableindicating import initiation from non-EU15 markets (STARTit) obtained from weighted firm-level regressions with weightsderived by propensity score matching. Regressions control for covariates included in first-stage probit regression. TableA.12 in the Appendix presents means of the outcome variables considered in this table.

38

Table 10: Direct imports and firm exit (H1)

i ii iii

DIMPit-0.026 -0.026 -0.026(0.000) (0.000) (0.000)

DSIZEit0.165 0.167 0.167

(0.000) (0.000) (0.000)

LN(V IMPst)0.007 0.001 0.001

(0.003) (0.787) (0.807)[0.018] [0.770] [0.786]

DIMPit× LN(V IMPst)-0.001 0.001(0.617) (0.478)[0.646] [0.532]

DSIZEit× LN(V IMPst)0.017 0.017

(0.000) (0.000)[0.016] [0.014]

LN(SALESst)0.020 0.019 0.019

(0.068) (0.075) (0.077)

Observations 119,058 119,058 119,058R-squared 0.067 0.071 0.071

Mean of dependent variable (exit dummy) 0.086 0.086 0.086

Notes: Firm-level regressions containing sector and year dummies. P-values referring to standard errors clustered at thesector level in parentheses. Square brackets present p-values referring to a wild bootstrap procedure with 1,000 replications.DIMPit is a dummy variable taking on unity if a firm imports consumer goods from outside EU15. DSIZEit is a dummyvariable taking on unity if a firm belongs to the lower 25th percentile of the size distribution. V IMPst are direct sector-levelconsumer goods imports from outside EU15. SALESst are sector-level sales.

Table 11: Direct imports and market-level adjustments (H5)

i ii iii iv

Unweighted regression Weighted regression

Dependent variableLN(PROD− LN(MARK− LN(PROD− LN(MARK−UCTIV ITY ) UPS) UCTIV ITY ) UPS)

LN(V IMPst)0.022 -0.007 0.016 0.001

(0.053) (0.090) (0.044) (0.896)[0.204] [0.154] [0.128] [0.894]

LN(ASSETSrs,t−1)0.022 0.013 0.019 0.024

(0.017) (0.036) (0.163) (0.035)

LN(MEANWAGErs,t−1)0.005 -0.036 -0.019 -0.036

(0.793) (0.000) (0.677) (0.061)

Observations 13,697 13,697 13,697 13,697R-squared 0.054 0.061 0.059 0.087

(Weighted) mean of dep. var. 12.815 0.244 12.864 0.282

Notes: Market-level regressions with market fixed effects and year dummies. Parentheses contain p-values referring tostandard errors clustered at the sector level; square brackets present p-values referring to a wild bootstrap procedure with1,000 replications. V IMPst are direct sector-level consumer goods imports from outside EU15. ASSETSrs,t−1 are thesum assets of all firms present in a market. MEANWAGErs,t−1 refers to the average wage in the market. Weights inweighted regressions are based on industry sales shares.

39

Table 12: Direct imports and market-level adjustments (H6, H7)

i ii iii iv v vi vii viii

Unweighted regression Weighted regression

Dependent variableLN(HER− LN(HER−

LN(NO. LN(MEAN LN(STD. FINDAHL LN(NO. LN(MEAN LN(STD. FINDAHLFIRMS) SALES) SALES) INDEX) FIRMS) SALES) SALES) INDEX)

LN(V IMPst)-0.034 0.022 0.018 0.034 -0.039 0.006 0.028 0.042(0.014) (0.311) (0.443) (0.001) (0.024) (0.792) (0.483) (0.054)[0.034] [0.508] [0.486] [0.014] [0.030] [0.884] [0.774] [0.134]

LN(ASSETSrst−1)0.131 0.154 0.216 -0.080 0.103 0.163 0.212 -0.044

(0.000) (0.000) (0.000) (0.000) (0.002) (0.000) (0.000) (0.059)

LN(MEAN WAGErst−1)-0.057 0.052 0.051 0.053 -0.035 0.059 0.061 0.063(0.008) (0.001) (0.200) (0.003) (0.203) (0.420) (0.688) (0.163)

Observations 13,697 13,697 11,873 13,697 13,697 13,697 11,873 13,697R-squared 0.070 0.104 0.051 0.030 0.064 0.113 0.079 0.025

(Weighted) mean of dep. var. 1.635 15.433 15.112 -1.114 2.203 15.885 15.869 -1.410

Notes: Market-level regressions with market fixed effects and year dummies. Parentheses contain p-values referring tostandard errors clustered at the sector level. Square brackets present p-values referring to a wild bootstrap procedure with1,000 replications. V IMPst are direct sector-level consumer goods imports from outside EU15. ASSETSrs,t−1 are thesum of assets of all firms present in a market. MEANWAGErs,t−1 refers to the average wage in the market. Weights inweighted regressions are based on industry sales shares.

40

Table 13: Indirect imports and firm exit (H1)

i ii iii

DIMPit-0.026 -0.026 -0.026(0.000) (0.000) (0.000)

DSIZEit0.165 0.167 0.167

(0.000) (0.000) (0.000)

LN(V IMPst)0.007 0.001 0.001

(0.003) (0.642) (0.663)[0.028] [0.696] [0.662]

DIMPit× LN(V IMPst)0.000 0.002

(0.965) (0.279)[0.952] [0.378]

DSIZEit× LN(V IMPst)0.016 0.016

(0.000) (0.000)[0.024] [0.022]

LN(V IMPOst )

-0.003 -0.005 -0.004(0.254) (0.107) (0.129)[0.494] [0.252] [0.292]

DIMPit× LN(V IMPOst )

-0.004 -0.004(0.216) (0.208)[0.216] [0.192]

DSIZEit× LN(V IMPOst )

0.002 0.002(0.808) (0.824)[0.792] [0.812]

LN(SALESst)0.023 0.023 0.023

(0.050) (0.062) (0.059)

Observations 119,058 119,058 119,058R-squared 0.067 0.071 0.071

Mean of dependent variable (exit dummy) 0.086 0.086 0.086

Notes: Firm-level regressions containing sector and year dummies. P-values referring to standard errors clustered at thesector level in parentheses. Square brackets present p-values referring to a wild bootstrap procedure with 1,000 replications.DIMPit is a dummy variable taking on unity if a firm imports consumer goods from outside EU15. DSIZEit is a dummyvariable taking on unity if a firm belongs to the lower 25th percentile of the size distribution. V IMPst and V IMPO

st aredirect and indirect sector-level consumer goods imports from outside EU15, respectively. SALESst are sector-level sales.

41

Table 14: Indirect imports and market-level adjustments (H5)

i ii iii iv

Unweighted regression Weighted regression

Dependent variableLN(PROD− LN(MARK− LN(PROD− LN(MARK−UCTIV ITY ) UPS) UCTIV ITY ) UPS)

LN(V IMPst)0.022 -0.007 0.016 0.001

(0.049) (0.098) (0.044) (0.887)[0.200] [0.200] [0.130] [0.896]

LN(V IMPOst )

-0.000 -0.001 0.000 -0.003(0.983) (0.875) (0.985) (0.801)[0.970] [0.864] [0.972] [0.966]

LN(ASSETSrs,t−1)0.022 0.013 0.019 0.024

(0.017) (0.036) (0.153) (0.034)

LN(MEAN WAGErs,t−1)0.005 -0.036 -0.019 -0.037

(0.791) (0.000) (0.671) (0.054)

Observations 13,697 13,697 13,697 13,697R-squared 0.054 0.061 0.059 0.087

(Weighted) mean of dep. var. 12.815 0.244 12.864 0.282

Notes: Market-level regressions with market fixed effects and year dummies. Parentheses contain p-values referring tostandard errors clustered at the sector level. Square brackets present p-values referring to a wild bootstrap procedure with1000 replications. V IMPst and V IMPO

st are direct and indirect sector-level consumer goods imports from outside EU15,respectively. ASSETSrs,t−1 are the sum of assets of all firms present in a market. MEANWAGErs,t−1 refers to theaverage wage in the market. Weights in weighted regressions are based on industry sales shares.

42

Table 15: Indirect imports and market-level adjustments (H6, H7)

i ii iii iv v vi vii viii

Unweighted regression Weighted regression

Dependent variableLN(HER− LN(HER−

LN(NO. LN(MEAN LN(STD. FINDAHL LN(NO. LN(MEAN LN(STD. FINDAHLFIRMS) SALES) SALES) INDEX) FIRMS) SALES) SALES) INDEX)

LN(V IMPst)-0.034 0.021 0.018 0.034 -0.039 0.004 0.026 0.042(0.014) (0.331) (0.435) (0.001) (0.024) (0.850) (0.529) (0.055)[0.034] [0.532] [0.478] [0.012] [0.026] [0.896] [0.736] [0.140]

LN(V IMPOst )

0.001 0.014 -0.001 -0.005 -0.003 0.048 0.063 0.002(0.955) (0.173) (0.972) (0.681) (0.707) (0.035) (0.257) (0.835)[1.000] [0.178] [1.000] [0.718] [0.712] [0.014] [0.350] [0.902]

LN(ASSETSrs,t−1)0.131 0.154 0.216 -0.080 0.104 0.159 0.206 -0.045

(0.000) (0.000) (0.000) (0.000) (0.001) (0.000) (0.000) (0.052)

LN(WAGErs,t−1)-0.057 0.053 0.051 0.053 -0.035 0.065 0.072 0.064(0.008) (0.001) (0.196) (0.003) (0.189) (0.369) (0.616) (0.160)

Observations 13,697 13,697 11,873 13,697 13,697 13,697 11,873 13,697R-squared 0.070 0.104 0.051 0.030 0.064 0.122 0.085 0.025

(Weighted) mean of dep. var. 1.635 15.433 15.112 -1.114 2.203 15.885 15.869 -1.410

Notes: Market-level regressions with market fixed effects and year dummies. Parentheses contain p-values referring tostandard errors clustered at the sector level; square brackets present p-values referring to a wild bootstrap procedure with1,000 replications. V IMPst and V IMPO

st are direct and indirect sector-level consumer goods imports from outside EU15,respectively. ASSETSrs,t−1 are the sum of assets of all firms present in a market. MEANWAGErs,t−1 refers to theaverage wage in the market. Weights in weighted regressions are based on industry sales shares.

43

A Appendix

This Appendix contains supplementary material including proofs for the hypotheses de-rived from the theoretical model, summary statistics, details on markup estimations, andadditional results in relation to propensity score matching.

A.1 Sufficient conditions for 0 < cS < cI < cD

To ensure that cI < cD, we assume that:

FI >L

4τ(w − t)2 . (29)

The condition cS < cI requires

FS >L

τ

(1

4(w − t) +

τFIL(w − t)

)2

. (30)

For 0 < cS it is sufficient to assume that for the most efficient firm, and thus for c = 0, itis strictly profitable to become a chain, which requires:

FS <L

4τ(cD + w − t)2 . (31)

A.2 Proof of Hypothesis 1

Using (2), (8) and (9), the expected zero-profit condition (12) can be rewritten as:∫ cS

0

[RL

4τ(cD + w − c− t)2 − (R− 1)FS − FI

]kck−1

ckMdc

+

∫ cI

cS

[L

4τ(cD + w − c− t)2 − FI

]kck−1

ckMdc

+

∫ cD

cI

L

4τ(cD − c)2 kc

k−1

ckMdc− FE = 0.

Applying the implicit function theorem and using the Leibniz rule, we obtain:

dcDdt

=(R− 1)cks

(w − t+ cD − k

k+1cS)

+ ckI(w − t+ cD − k

k+1cI)

(R− 1)cks(w − t+ cD − k

k+1cs)

+ (w − t)ckI +ck+1D

k+1

> 0, (32)

since w > t, and cD > kk+1

cI >kk+1

cS due to cD > cI > cs and k > 0.

44

A.3 Proof of Hypothesis 2

From (10), we obtain:

dcIdt

=dcDdt−(

1

2+

2τFIL(w − t)2

)=

dcDdt− 1− cD − cI

w − t.

Substituting for dcDdt

, we obtain:

dcIdt

=Num

Den< 0,

since:

Den = (R− 1)cks

(w − t+ cD −

k

k + 1cs

)+ (w − t)ckI +

ck+1D

k + 1> 0,

and:

Num = (R− 1)cks

(w − t+ cD −

k

k + 1cS

)+ ckI

(w − t+ cD −

k

k + 1cI

)−(

1 +cD − cIw − t

)[(R− 1)cks

(w − t+ cD −

k

k + 1cs

)+ (w − t)ckI +

ck+1D

k + 1

]= −c

k+1D − ck+1

I

k + 1− cD − cI

w − tck+1D

k + 1

−(

1 +cD − cIw − t

)[(R− 1)cks

(w − t+ cD −

k

k + 1cs

)]< 0.

A.4 Proof of Hypothesis 3

Differentiating (6) and (8) with respect to t and using (32), we obtain:

dqD

dt=

L

2τ

dcDdt

> 0 anddπD

dt=

L

2τ(cD − c)

dcDdt

> 0.

Differentiating (7) and (9) with respect to t and using dcDdt

< 1, we have:

dqI

dt=

L

2τ

[dcDdt− 1

]< 0 and

dπI

dt=

L

2τ(cD + w − t− c)

[dcDdt− 1

]< 0.

The result on markups and profits follows immediately, as markups are proportional tooutput.

45

A.5 Proof of Hypothesis 4

From (11) we have:dcsdt

=dcDdt− 1.

Note that dcDdt

< 1, if:

ck+1D + kck+1

I > (k + 1) ckIcD, (33)

where (33) is satisfied with equality for cD = cI , and the LHS increases faster with cDthan the RHS. This proves dcs

dt< 0.

A.6 Proof of Hypothesis 6

dn

dt=

τ

(cD + w − p)2

(dp

dt− dcD

dt

). (34)

Hence:

sign

{dn

dt

}= sign

{dp

dt− dcD

dt

}.

After substituting fordpdt

and dcDdt

, we have:

sign

{dn

dt

}= sign

{(−1

2(k + 1)

)dcDdt

+1

2

ckIckD

+k(w − t)

2

ckIckD

[1

cD

dcDdt− 1

cI

dcIdt

]}. (35)

For FI equal to its lower bound (FI = L4τ

(w − t)2), we have cD = cI ,dcIdt

= 0, and dcDdt

= 1.Therefore:

sign

{dn

dt

}= sign

{k

2(k + 1)+k(w − t)

2cD

}> 0.

A.7 Proof of Hypothesis 7

The average sales volume of a shop is given by:

q = q(p) =L

τ

(cD

2(k + 1)+

(w − t)2

ckIckD

), (36)

and the derivative with respect to t is:

dq

dt=

L

2τ

((w − t)c

k−1I

ck+1D

(cDdcIdt− cI

dcDdt

)− ckIckD

+1

k + 1

dcDdt

). (37)

For FI at its lower bound, we have cD = cI ,dcIdt

= 0, and dcDdt

= 1. Using these values in(37), we obtain:

dq

dt= − L

2τ

((w − t) 1

cD+

k

k + 1

)< 0.

46

The variance of retail sales is given by:

σ2q =

L2

4τ 2

{kc2

D

(k + 2)(k + 1)2

+

((w − t)2

[1− ckI

ckD

]+

2k(cD − cI)(w − t)(k + 1)

)ckIckD

}. (38)

Evaluating the derivative at the lower bound of FI we obtain:

dσ2q

dt=

L2

4τ 2

{(w − t)2 k

cD+

2kcD(k + 2)(k + 1)2

}> 0.

47

A.8 Summary statistics and sector definitions

Tables A.1 and A.2 contain summary statistics for the main variables used in the empiricalanalysis conducted at the firm level and market level, respectively. Table A.3 presentsthe retail sectors used in the empirical analysis.

Table A.1: Summary statistics - firm-level analyses

obs mean sd p10 p25 p50 p75 p90

LN(SALESit) 92,021 15.33 1.13 14.09 14.62 15.21 15.92 16.76LN(PROFITSit) 79,251 12.60 1.13 11.23 12.01 12.66 13.25 13.85LN(MARKUPit) 75,370 0.15 0.24 -0.15 -0.01 0.14 0.30 0.46CHAIN DUMMY 88,319 0.10 0.30NEW-SHOP DUMMY 87,437 0.03 0.16LN(EMPLOY EESit−1) 92,021 1.18 1.03 0.00 0.00 1.10 1.79 2.48LN(PRODUCTIV ITYit−1) 90,900 12.75 0.48 12.23 12.51 12.74 13.03 13.33DIMPit 92,021 0.09 0.29EXIT DUMMY 119,058 0.09 0.28LN(V IMPst) 119,058 18.49 2.25 14.99 16.98 19.09 20.41 20.58LN(V IMPO

st ) 119,058 17.52 0.90 16.28 17.19 17.60 18.02 18.50

Table A.2: Summary statistics - market-level analyses

obs mean sd p10 p25 p50 p75 p90

LN(FIRMSrst) 13,697 1.63 1.08 0.00 0.69 1.61 2.30 3.09LN(MEAN SALESrst) 13,697 15.43 1.03 14.42 14.79 15.24 15.84 16.71LN(STD. SALESrst) 11,873 15.11 1.37 13.63 14.27 14.97 15.78 16.94LN(HERFINDAHLINDEXrst) 13,697 -1.11 0.86 -2.30 -1.68 -1.06 -0.44 0.00LN(PRODUCTIV ITYrst) 13,697 12.82 0.28 12.51 12.65 12.79 12.97 13.16LN(MARKUPrst) 13,697 0.24 0.19 0.01 0.12 0.23 0.36 0.49LN(ASSETSrst−1) 13,697 16.34 1.59 14.40 15.24 16.22 17.38 18.51LN(WAGErst−1) 13,697 11.53 0.46 10.97 11.26 11.55 11.83 12.08LN(V IMPst) 13,697 17.44 2.18 14.60 15.74 17.60 18.94 20.41LN(V IMPO

st ) 13,697 17.24 1.10 15.73 16.57 17.44 17.92 18.50

48

Table A.3: Sector classification

Nace codes Sector description

52 Retail trade, except of motor vehicles and motorcycles; repair of personal and household goods52.1 Retail sale in non-specialized stores52.11 Retail sale in non-specialized stores with food, beverages or tobacco predominating52.11.10 Grocer’s shops52.11.20 All-night shops52.11.30 Supermarkets52.12 Other retail sale in non-specialized stores52.12.10 Variety stores52.12.20 Department stores52.4 Other retail sale of new goods in specialized stores52.41 Retail sale of textiles52.41.00 Retail sale of textiles52.42 Retail sale of clothing52.42.10 Retail sale of ladies’ clothing52.42.20 Retail sale of men’s clothing52.42.30 Retail sale of men’s and ladies’ clothing52.42.40 Retail sale of baby articles and children’s clothing52.43 Retail sale of footwear and leather goods52.43.10 Retail sale of footwear52.43.20 Retail sale of leather goods52.44 Retail sale of furniture, lighting equipment and household articles n.e.c.52.44.10 Retail sale of furniture52.44.20 Retail sale of carpets52.44.30 Retail sale of furnishing fabrics52.44.40 Retail sale of kitchen utensils, glass and china52.44.50 Retail sale of articles for lighting52.45 Retail sale of electrical household appliances and radio and television goods52.45.10 Retail sale of electric household appliances52.45.20 Retail sale of radio and television goods52.45.30 Retail sale of records, CDs, cassettes, etc.52.45.40 Retail sale of musical instruments52.46 Retail sale of hardware, paints and glass52.46.10 Retail sale of hardware52.46.20 Retail sale of building materials52.46.30 Retail sale of paints and wallpaper52.48 Other retail sale in specialized stores52.48.05 Retail sale of watches and clocks52.48.10 Retail sale of watches, clocks and jewellery52.48.15 Retail sale of jewellery52.48.20 Retail sale of glasses52.48.25 Retail sale of photographic equipment52.48.30 Gift shops52.48.35 Art shops and galleries52.48.40 Retail sale of stamps and coins52.48.45 Retail sale of sports goods52.48.50 Retail sale of toys and games

49

52.48.55 Retail sale of pleasure boats and parts thereof52.48.60 Retail sale of bicycles and mopeds52.48.65 Retail sale of computers, office machinery and standard software52.48.70 Retail sale of telecommunications equipment52.48.75 Florist’s shops52.48.80 Retail sale of plants and seeds52.48.85 Retail sale of pet animals52.48.90 Retail sale of fuel oils and solid fuels for households52.48.95 Sex shops52.48.99 Retail sale of other goods

Notes: Sector definition used in paper: NACE four-digit except for sector 5248, where a more detailed sector definitionis used. Usually, the six-digit sector codes are used; the exceptions are the aggregation of sectors 524805-524815, 524865-524870, and 524875-524880.

50

A.9 Markup estimation

As noted in the main text, we obtain a measure for firm-level markups following themethodology of De Loecker and Warzynski (2012), which builds on insights from Hall(1988) and relies on fairly modest assumptions; specifically, the approach rests on theassumptions that firms minimize costs and that there is at least one variable input thatis free of adjustment costs. The methodology does not depend on a particular type ofcompetition or functional form of demand, and it accommodates a variety of (static) pricesetting models. Specifically, the markup of firm i in period t, µit, can be obtained fromthe following relationship:

µit =εXit

αXit/exp(uit)

, (39)

where αXit is the share of input X in total output of firm i, uit is a correction factor forunobserved shocks to production, and εXit represents the output elasticity of input X.While αit is directly observable in our data, we require estimates of εXit and uit. FollowingDe Loecker and Warzynski (2012), we obtain both objects by estimating a translog valueadded production function, as explained further below. Note that we use the outputelasticity of labor to recover firm-level markups. Given the highly flexible labor market inDenmark, considering labor as the variable input that is free of adjustment costs appearsreasonable.

We estimate the production function using a control function approach to deal withsimultaneity problems related to unobserved productivity shocks. Specifically, we followthe insights of Ackerberg, Caves and Frazer (2015), while relying on material inputsto proxy for unobserved productivity, as suggested by Levinsohn and Petrin (2003).55

The approach involves estimating the output elasticities using GMM techniques withthe identifying assumption that productivity follows a first order Markov process. Theadjustment factor uit is obtained from the first stage of the estimation algorithm, as alsorecently documented by Brandt et al. (2017).

The implementation of the estimation approach requires firm-level variables on valueadded, labor, capital, and materials, which we obtain from the FIRM data set. We deflatevalue added, material inputs, and the capital stock using the consumer price index, amaterial goods price deflator, and a capital goods price deflator, respectively. The wagebill used to compute markups is also reported in the FIRM data set.56

Table A.4 presents estimates of the median output elasticities and markups by aggre-gated retail sector. Note that we estimate the production function separately for each ofthe sectors presented in the table. Additional summary statistics for markups are pre-sented in Table A.1. Median markups of retail firms amount to 15% (the mean equals20%), while the tables display a considerable degree of heterogeneity of markups acrossboth sectors and firms.

55 As suggested by De Loecker and Warynski, we control for a firm’s import and export status in theestimation approach.

56 When computing the markups, we trim the first and last percentiles of uit, αXit , and µit to account

for some extreme observations in the data.

51

Table A.4: Median output elasticities and markups by sector

coeff. coeff. obser-labor capital markup vations

Retail sale of textiles, apparel, footwear, 0.739 0.112 1.095 22,994and leather goods (0.118) (0.058) (0.297) 22,994Retail sale of furniture, lighting equipment, 0.798 0.106 1.127 7,978and household articles n.e.c. (0.122) (0.059) (0.318) 7,978Retail sale of electrical household app- 0.764 0.113 1.063 5,290liances and radio and television goods (0.107) (0.053) (0.285) 5,290

Retail sale of hardware, paints and glass0.800 0.091 1.145 5,432

(0.132) (0.052) (0.314) 5,432Retail sale in non-specialized stores 0.780 0.147 1.167 18,196(supermarkets, department stores) (0.091) (0.066) (0.275) 18,196

Other retail sale in specialized stores0.787 0.116 1.203 31,848

(0.139) (0.058) (0.321) 31,848

All sectors median 1.150mean 1.199

Notes: Standard deviation in parentheses.

52

A.10 Additional PSM results

In this section, we present some additional information about the propensity score match-ing (PSM) approach. First, we present estimation results for the probit model estimatedin the first stage of the matching approach in Table A.5. The table contains results forthe outcome variables sales, profits, and markups referring to the period p = 0. Thecovariates indicate plausible relationships. Larger firms (measured by total assets) havea higher probability of starting to import. The same holds for more productive firms asindicated by the positive coefficients of the (de-meaned and squared) productivity terms.Moreover, firms with larger sales growth are more likely to become importers. In TableA.6, we present results related to the balancing properties after applying radius match-ing with a caliper of 0.001. The results show that the matching strategy indeed leadsto very similar means of the covariates. In particular, the p-values indicate that we cannever reject the null hypothesis of equal means and we obtain very low numbers for thestandardized bias. The balancing properties also hold when considering matched sam-ples with outcomes in later periods. Table A.7 shows that the mean standardized biasis always well below three also in later periods (see Caliendo and Kopening (2008) for adiscussion of the size of the standardized bias). This is also true for the alternative match-ing approaches depicted in the table. ATTs for these alternative matching approachesare presented in consecutive tables. First, we add the lagged (log-) level of the outcomevariable to the probit model. We present these results in Table A.8, while noting thatImbens and Wooldridge (2009) point out that the inclusion of lagged outcomes may implydifferent identifying restrictions on the data. Second, Table A.9 contains results referringto exact matching by year and NACE four-digit sector which is a more restrictive match-ing approach. Third, results in Table A.10 are obtained when applying matching witha Gaussian kernel (bandwidth of 0.001). Overall, results are qualitatively similar acrossthese tables. Finally, Tables A.11 and A.12 present means of the dependent variablesconsidered in Tables 6 and 9 in the main text.

53

Table A.5: Probit regressions for samples with outcomes at p=0

SALES PROFITS MARKUPS

LN(PRODUCTIV ITYit−1) 0.016 -0.007 0.074(0.594) (0.856) (0.076)

LN(EMPLOY EESit−1) 0.033 0.037 0.061(0.179) (0.191) (0.034)

SALESGROWTHit−1 0.106 0.091 0.081(0.000) (0.003) (0.012)

WAGESHAREit−1 -0.269 -0.258 -0.124(0.093) (0.201) (0.556)

LN(ASSETSit−1) 0.155 0.150 0.158(0.000) (0.000) (0.000)

LN(PRODUCTIV ITYit−1)2 0.084 0.118 0.118(0.000) (0.001) (0.010)

Observations 77401 57362 60589Pseudo R-square 0.0694 0.0706 0.0752

Notes: Table presents results from probit regressions with a dummy variable indicating import initiation from non-EU15 markets as dependent variable. Regressions contain NACE four-digit industry and year dummies. P-values inparentheses. PRODUCTIV ITYit−1 refers to value added per employee; EMPLOY EESit−1 to number of employees;SALESGROWTHit−1 to growth in domestic sales; WAGESHAREit−1 to the share of wages in total sales; ASSETSit−1

to total assets; and LN(PRODUCTIV ITYit−1)2 to the square of (demeaned) log productivity. LN(·) indicates log trans-formation.

54

Table A.6: Balancing properties after radius matching for samples with outcomes at p=0

Mean standardizedTreated Control bias p-value

Outcome Variable: LN(SALESit)LN(PRODUCTIV ITYit−1) 12.797 12.794 0.6 0.881LN(EMPLOY EESit−1) 1.305 1.297 0.7 0.862SALESGROWTHit−1 0.132 0.136 -0.9 0.821WAGESHAREit−1 0.178 0.177 1.4 0.709LN(ASSETSit−1) 14.748 14.739 0.8 0.840LN(PRODUCTIV ITYit−1)2 0.243 0.254 -2.0 0.604

Outcome Variable: LN(PROFITSit)LN(PRODUCTIV ITYit−1) 12.858 12.853 1.4 0.763LN(EMPLOY EESit−1) 1.249 1.253 -0.4 0.937SALESGROWTHit−1 0.127 0.128 -0.2 0.957WAGESHAREit−1 0.171 0.170 1.1 0.801LN(ASSETSit−1) 14.713 14.712 0.1 0.980LN(PRODUCTIV ITYit−1)2 0.178 0.181 -0.8 0.866

Outcome Variable: LN(MARKUPSit)LN(PRODUCTIV ITYit−1) 12.818 12.813 1.3 0.764LN(EMPLOY EESit−1) 1.472 1.477 -0.5 0.916SALESGROWTHit−1 0.112 0.117 -1.1 0.811WAGESHAREit−1 0.186 0.186 0.6 0.884LN(ASSETSit−1) 14.903 14.902 0.1 0.985LN(PRODUCTIV ITYit−1)2 0.172 0.177 -1.9 0.693

Notes: Table presents information about balancing properties of covariates after radius matching with a caliper of0.001 for outcome variables sales, profits, and markups. PRODUCTIV ITYit−1 refers to value added per employee;EMPLOY EESit−1 to number of employees; SALESGROWTHit−1 to growth in domestic sales; WAGESHAREit−1

to the share of wages in total sales; ASSETSit−1 to total assets; and LN(PRODUCTIV ITYit−1)2 to the square of(demeaned) log productivity. LN(·) indicates log transformation.

55

Table A.7: Mean standardized bias after applying various matching approaches across allconsidered time horizons

p p+1 p+2 p+3

Baseline matching approachSales 1.096 0.906 0.906 0.906Profits 0.667 0.974 0.974 0.974Markups 0.914 1.036 1.036 1.036

Lagged level of outcome variable includedSales 1.151 0.570 1.549 1.099Profits 1.005 1.529 0.803 0.994Markups 0.748 1.015 1.035 1.265

Exact matching by year and sectorSales 0.827 0.845 1.722 0.994Profits 0.850 0.816 1.457 0.870Markups 1.177 1.086 1.733 1.580

Matching using a Gaussian kernelSales 0.556 0.650 1.105 0.989Profits 0.719 2.154 1.912 1.017Markups 0.771 0.828 1.266 0.859

Notes: Table presents information about the mean of the standardized bias of covariates in the probit model after applyingthe matching approach.

56

Table A.8: Direct importing and firm performance - matching (H3) with lagged outcomesin first-stage probit regression

p p+1 p+2 p+3

Domestic Sales

LN(SALES)0.098 0.092 0.121 0.118

(0.001) (0.009) (0.002) (0.008)

∆LN(SALES)0.087 0.084 0.101 0.105

(0.000) (0.000) (0.000) (0.000)

Number of treated 1,354 1,154 943 943Observations 68,871 54,635 41,479 31,357

Profits

LN(PROFITS)0.082 0.103 0.099 0.138

(0.021) (0.010) (0.033) (0.009)

∆LN(PROFITS)0.072 0.095 0.090 0.122

(0.012) (0.006) (0.033) (0.015)

Number of treated 1,040 831 647 495Observations 50,464 37,131 26,649 18,571

Markups

LN(MARKUPS)0.016 0.025 0.023 0.024

(0.028) (0.003) (0.016) (0.017)

∆LN(MARKUPS)0.017 0.022 0.024 0.023

(0.014) (0.007) (0.017) (0.033)

Number of treated 1,076 876 679 522Observations 52,418 39,312 28,075 19,974

Notes: P-values referring to standard errors clustered at the firm level in parentheses. Coefficients refer to a dummyvariable indicating import initiation from non-EU15 markets (STARTit) obtained from weighted firm-level regressions withweights derived by propensity score matching. Regressions control for covariates included in first-stage probit regression.∆ indicates that changes in the outcome variable are considered. Note that these differences are computed relative topre-treatment status (i.e. relative to p-1.)

57

Table A.9: Direct importing and firm performance - matching (H3) by year and sector

p p+1 p+2 p+3

Domestic Sales

LN(SALES)0.085 0.085 0.103 0.116

(0.000) (0.000) (0.000) (0.000)

∆LN(SALES)0.089 0.092 0.109 0.108

(0.000) (0.000) (0.000) (0.000)

Number of treated 1,331 1,127 910 734Observations 52,961 40,638 30,093 20,852

Profits

LN(PROFITS)0.067 0.076 0.087 0.094

(0.013) (0.015) (0.020) (0.024)

∆LN(PROFITS)0.099 0.103 0.116 0.151

(0.000) (0.002) (0.004) (0.002)

Number of treated 1,018 807 624 471Observations 38,517 27,529 18,774 11,441

Markups

LN(MARKUPS)0.023 0.027 0.025 0.025

(0.002) (0.001) (0.009) (0.016)

∆LN(MARKUPS)0.014 0.016 0.016 0.011

(0.045) (0.049) (0.102) (0.315)

Number of treated 1,043 848 646 495Observations 38,146 26,531 17,658 11,740

Notes: P-values referring to standard errors clustered at the firm level in parentheses. Coefficients refer to a dummyvariable indicating import initiation from non-EU15 markets (STARTit) obtained from weighted firm-level regressions withweights derived by propensity score matching. Regressions control for covariates included in first-stage probit regression.∆ indicates that changes in the outcome variable are considered. Note that these differences are computed relative topre-treatment status (i.e. relative to p-1.)

58

Table A.10: Direct importing and firm performance - matching (H3) using a Gaussiankernel

p p+1 p+2 p+3

Domestic Sales

LN(SALES)0.081 0.082 0.102 0.115

(0.000) (0.000) (0.000) (0.000)

∆LN(SALES)0.088 0.092 0.101 0.109

(0.000) (0.000) (0.000) (0.000)

Number of treated 1,368 1,165 953 775Observations 77,384 62,505 49,951 39,470

Profits

LN(PROFITS)0.069 0.101 0.092 0.098

(0.009) (0.001) (0.011) (0.013)

∆LN(PROFITS)0.091 0.118 0.119 0.159

(0.001) (0.000) (0.002) (0.001)

Number of treated 1,049 841 657 504Observations 57,348 43,375 32,794 24,518

Markups

LN(MARKUPS)0.022 0.030 0.027 0.029

(0.002) (0.000) (0.003) (0.002)

∆LN(MARKUPS)0.012 0.018 0.015 0.015

(0.069) (0.016) (0.103) (0.149)

Number of treated 1,088 888 691 541Observations 60,567 46,524 35,930 27,504

Notes: P-values referring to standard errors clustered at the firm level in parentheses. Coefficients refer to a dummyvariable indicating import initiation from non-EU15 markets (STARTit) obtained from weighted firm-level regressions withweights derived by propensity score matching. Regressions control for covariates included in first-stage probit regression.∆ indicates that changes in the outcome variable are considered. Note that these differences are computed relative topre-treatment status (i.e. relative to p-1.)

Table A.11: Mean of outcome variables in Table 6

p p+1 p+2 p+3

LN(SALES) 15.383 15.395 15.418 15.480LN(SALES) -0.038 -0.041 -0.023 0.001

Observations 69,025 54,910 41,469 31,256

ProfitsLN(PROFITS) 12.774 12.836 12.894 12.992LN(PROFITS) -0.001 0.036 0.095 0.170

Observations 50,329 36,954 26,765 18,496Mark-ups

LN(MARKUPS) 0.165 0.171 0.175 0.178LN(MARKUPS) -0.029 -0.028 -0.016 -0.009

Observations 52,598 38,976 27,854 19,863

Notes: Mean refers to a weighted average with weights obtained from the matching approach.

59

Table A.12: Mean of outcome variables in Table 9

p-1 p p+1 p+2 p+3

CHAIN DUMMY 0.016 0.032 0.055 0.069Observations 58,318 44,968 33,323 24,461

NEW-SHOP DUMMY 0.018 0.019 0.032 0.044 0.052Observations 64,474 66,154 51,284 38,442 28,877

Notes: Mean refers to a weighted average with weights obtained from the matching approach.

60